From 1493aed81d9215eb3ed9f172ce266d6e95552e54 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Tue, 26 May 2026 22:19:38 +0800
Subject: [PATCH 01/14] chore: inspect ai generation flow


From c9fea718215e2c204240005968057198932fa1f3 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Tue, 26 May 2026 23:10:41 +0800
Subject: [PATCH 02/14] feat: add ark seedance video generation

---
 .env.example                       |   4 +
 config/config.go                   |   1 +
 handler/ai.go                      |  38 ++++-
 handler/media_reference.go         | 172 +++++++++++++++++++++++
 router/router.go                   |   4 +
 service/settings.go                |   2 +-
 web/src/services/api/image.ts      |   8 +-
 web/src/services/api/video.ts      | 218 +++++++++++++++++++++++++++--
 web/src/stores/use-config-store.ts |   2 +-
 web/src/types/media.ts             |   7 +
 10 files changed, 436 insertions(+), 20 deletions(-)
 create mode 100644 handler/media_reference.go
 create mode 100644 web/src/types/media.ts

diff --git a/.env.example b/.env.example
index 4708dad..6b7ae19 100644
--- a/.env.example
+++ b/.env.example
@@ -9,6 +9,10 @@ JWT_EXPIRE_HOURS=168
 # 后端监听端口
 PORT=8080
 
+# 公开访问地址，用于把本地上传的 Seedance 参考图/视频暴露给火山方舟拉取。
+# 线上部署时填写站点根地址，例如：https://your-domain.example.com
+# PUBLIC_BASE_URL=https://your-domain.example.com
+
 # 前端开发代理默认使用 http://127.0.0.1:8080，如后端开发端口不同，启动前端时可单独设置 API_BASE_URL。
 # API_BASE_URL=http://127.0.0.1:8080
 
diff --git a/config/config.go b/config/config.go
index f91f51f..95976b0 100644
--- a/config/config.go
+++ b/config/config.go
@@ -17,6 +17,7 @@ type Config struct {
 	JWTExpireHours      int    `env:"JWT_EXPIRE_HOURS" envDefault:"168"`
 	StorageDriver       string `env:"STORAGE_DRIVER" envDefault:"sqlite"`
 	DatabaseDSN         string `env:"DATABASE_DSN" envDefault:"data/infinite-canvas.db"`
+	PublicBaseURL       string `env:"PUBLIC_BASE_URL"`
 	LinuxDoAuthorizeURL string `env:"LINUX_DO_AUTHORIZE_URL" envDefault:"https://connect.linux.do/oauth2/authorize"`
 	LinuxDoTokenURL     string `env:"LINUX_DO_TOKEN_URL" envDefault:"https://connect.linux.do/oauth2/token"`
 	LinuxDoUserInfoURL  string `env:"LINUX_DO_USERINFO_URL" envDefault:"https://connect.linux.do/api/user"`
diff --git a/handler/ai.go b/handler/ai.go
index f2c3389..045936c 100644
--- a/handler/ai.go
+++ b/handler/ai.go
@@ -49,6 +49,7 @@ func proxyAIGetRequest(w http.ResponseWriter, r *http.Request, path string) {
 		Fail(w, "AI 接口请求失败")
 		return
 	}
+	path = resolveAIProxyPath(channel.BaseURL, modelName, path)
 	request, err := http.NewRequest(http.MethodGet, service.BuildModelChannelURL(channel, path), nil)
 	if err != nil {
 		Fail(w, "AI 接口请求失败")
@@ -83,6 +84,7 @@ func proxyAIRequest(w http.ResponseWriter, r *http.Request, path string) {
 		Fail(w, "AI 接口请求失败")
 		return
 	}
+	path = resolveAIProxyPath(channel.BaseURL, modelName, path)
 	request, err := http.NewRequest(http.MethodPost, service.BuildModelChannelURL(channel, path), bytes.NewReader(body))
 	if err != nil {
 		log.Printf("AI proxy build request failed: url=%s err=%v", service.BuildModelChannelURL(channel, path), err)
@@ -117,12 +119,12 @@ func copyAIResponse(w http.ResponseWriter, request *http.Request, onFailure func
 	defer response.Body.Close()
 
 	if response.StatusCode >= http.StatusBadRequest {
-		payload, _ := io.ReadAll(io.LimitReader(response.Body, 4096))
-		log.Printf("AI upstream error: url=%s status=%d body=%s", request.URL.String(), response.StatusCode, strings.TrimSpace(string(payload)))
+		_, _ = io.Copy(io.Discard, io.LimitReader(response.Body, 4096))
+		log.Printf("AI upstream error: url=%s status=%d", request.URL.String(), response.StatusCode)
 		if onFailure != nil {
 			onFailure()
 		}
-		Fail(w, "AI 接口请求失败")
+		Fail(w, aiStatusMessage(response.StatusCode))
 		return
 	}
 
@@ -207,6 +209,36 @@ func readAIRequestCount(body []byte, contentType string) int {
 
 var errMissingModel = &aiError{"缺少模型名称"}
 
+func resolveAIProxyPath(baseURL string, modelName string, path string) string {
+	if !isArkSeedanceVideo(baseURL, modelName) {
+		return path
+	}
+	if path == "/videos" {
+		return "/contents/generations/tasks"
+	}
+	if strings.HasPrefix(path, "/videos/") && !strings.HasSuffix(path, "/content") {
+		return "/contents/generations/tasks/" + strings.TrimPrefix(path, "/videos/")
+	}
+	return path
+}
+
+func isArkSeedanceVideo(baseURL string, modelName string) bool {
+	base := strings.ToLower(baseURL)
+	model := strings.ToLower(modelName)
+	return strings.Contains(model, "seedance") || strings.Contains(model, "doubao-seedance") || strings.Contains(base, "/api/plan/v3")
+}
+
+func aiStatusMessage(statusCode int) string {
+	switch statusCode {
+	case http.StatusUnauthorized, http.StatusForbidden:
+		return "AI 接口鉴权失败，请检查 API Key、套餐权限或模型权限"
+	case http.StatusTooManyRequests:
+		return "AI 接口限流或额度不足，请稍后重试或检查额度"
+	default:
+		return "AI 接口请求失败"
+	}
+}
+
 type aiError struct {
 	message string
 }
diff --git a/handler/media_reference.go b/handler/media_reference.go
new file mode 100644
index 0000000..d080f97
--- /dev/null
+++ b/handler/media_reference.go
@@ -0,0 +1,172 @@
+package handler
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/basketikun/infinite-canvas/config"
+	"github.com/google/uuid"
+)
+
+const referenceMediaMaxBytes = 80 << 20
+
+type referenceMediaUploadResult struct {
+	ID       string `json:"id"`
+	URL      string `json:"url"`
+	MimeType string `json:"mimeType"`
+	Bytes    int64  `json:"bytes"`
+}
+
+func UploadReferenceMedia(w http.ResponseWriter, r *http.Request) {
+	publicBaseURL := strings.TrimRight(strings.TrimSpace(config.Cfg.PublicBaseURL), "/")
+	if publicBaseURL == "" {
+		Fail(w, "未配置 PUBLIC_BASE_URL，无法把本地参考视频提供给火山方舟访问")
+		return
+	}
+	r.Body = http.MaxBytesReader(w, r.Body, referenceMediaMaxBytes+1)
+	if err := r.ParseMultipartForm(referenceMediaMaxBytes); err != nil {
+		Fail(w, "参考素材过大或上传格式不正确")
+		return
+	}
+	if r.MultipartForm != nil {
+		defer r.MultipartForm.RemoveAll()
+	}
+	file, header, err := r.FormFile("file")
+	if err != nil {
+		Fail(w, "请上传参考图片或视频")
+		return
+	}
+	defer file.Close()
+
+	mimeType, ext, ok := normalizeReferenceMediaType(header.Header.Get("Content-Type"), filepath.Ext(header.Filename))
+	if !ok {
+		Fail(w, "参考素材格式不支持，请使用 jpeg/png/webp/bmp/gif/heic/heif 图片或 mp4/mov 视频")
+		return
+	}
+	if err := os.MkdirAll(referenceMediaDir(), 0o755); err != nil {
+		Fail(w, "参考素材保存失败")
+		return
+	}
+	id := uuid.NewString() + ext
+	targetPath := filepath.Join(referenceMediaDir(), id)
+	target, err := os.OpenFile(targetPath, os.O_WRONLY|os.O_CREATE|os.O_EXCL, 0o644)
+	if err != nil {
+		Fail(w, "参考素材保存失败")
+		return
+	}
+	bytes, copyErr := io.Copy(target, file)
+	closeErr := target.Close()
+	if copyErr != nil || closeErr != nil {
+		_ = os.Remove(targetPath)
+		Fail(w, "参考素材保存失败")
+		return
+	}
+	if bytes <= 0 || bytes > referenceMediaMaxBytes {
+		_ = os.Remove(targetPath)
+		Fail(w, "参考素材为空或超过大小限制")
+		return
+	}
+	OK(w, referenceMediaUploadResult{
+		ID:       id,
+		URL:      fmt.Sprintf("%s/api/media/references/%s", publicBaseURL, id),
+		MimeType: mimeType,
+		Bytes:    bytes,
+	})
+}
+
+func ReferenceMedia(w http.ResponseWriter, r *http.Request, id string) {
+	if id == "" || id != filepath.Base(id) || strings.Contains(id, "..") {
+		http.NotFound(w, r)
+		return
+	}
+	path := filepath.Join(referenceMediaDir(), id)
+	file, err := os.Open(path)
+	if err != nil {
+		http.NotFound(w, r)
+		return
+	}
+	defer file.Close()
+	info, err := file.Stat()
+	if err != nil || info.IsDir() {
+		http.NotFound(w, r)
+		return
+	}
+	if mimeType := mimeTypeByReferenceMediaExt(filepath.Ext(id)); mimeType != "" {
+		w.Header().Set("Content-Type", mimeType)
+	}
+	w.Header().Set("Cache-Control", "public, max-age=86400")
+	http.ServeContent(w, r, id, info.ModTime(), file)
+}
+
+func referenceMediaDir() string {
+	return filepath.Join("data", "reference-media")
+}
+
+func normalizeReferenceMediaType(contentType string, ext string) (string, string, bool) {
+	contentType = strings.ToLower(strings.TrimSpace(strings.Split(contentType, ";")[0]))
+	ext = strings.ToLower(strings.TrimSpace(ext))
+	if contentType == "" || contentType == "application/octet-stream" {
+		contentType = mimeTypeByReferenceMediaExt(ext)
+	}
+	if fixedExt := referenceMediaExtByMimeType(contentType); fixedExt != "" {
+		return contentType, fixedExt, true
+	}
+	if mimeType := mimeTypeByReferenceMediaExt(ext); mimeType != "" {
+		return mimeType, ext, true
+	}
+	return "", "", false
+}
+
+func referenceMediaExtByMimeType(mimeType string) string {
+	switch strings.ToLower(mimeType) {
+	case "image/jpeg", "image/jpg":
+		return ".jpg"
+	case "image/png":
+		return ".png"
+	case "image/webp":
+		return ".webp"
+	case "image/bmp":
+		return ".bmp"
+	case "image/gif":
+		return ".gif"
+	case "image/heic":
+		return ".heic"
+	case "image/heif":
+		return ".heif"
+	case "video/mp4":
+		return ".mp4"
+	case "video/quicktime", "video/mov":
+		return ".mov"
+	default:
+		return ""
+	}
+}
+
+func mimeTypeByReferenceMediaExt(ext string) string {
+	switch strings.ToLower(ext) {
+	case ".jpg", ".jpeg":
+		return "image/jpeg"
+	case ".png":
+		return "image/png"
+	case ".webp":
+		return "image/webp"
+	case ".bmp":
+		return "image/bmp"
+	case ".gif":
+		return "image/gif"
+	case ".heic":
+		return "image/heic"
+	case ".heif":
+		return "image/heif"
+	case ".mp4":
+		return "video/mp4"
+	case ".mov":
+		return "video/quicktime"
+	default:
+		return ""
+	}
+}
diff --git a/router/router.go b/router/router.go
index acf5bb7..81525d5 100644
--- a/router/router.go
+++ b/router/router.go
@@ -22,11 +22,15 @@ func New() *gin.Engine {
 	api.GET("/auth/linux-do/callback", gin.WrapF(handler.LinuxDoCallback))
 	api.GET("/auth/me", middleware.OptionalAuth, gin.WrapF(handler.CurrentUser))
 	api.GET("/settings", gin.WrapF(handler.Settings))
+	api.GET("/media/references/:id", func(c *gin.Context) {
+		handler.ReferenceMedia(c.Writer, c.Request, c.Param("id"))
+	})
 	v1 := api.Group("/v1", middleware.UserAuth)
 	v1.POST("/images/generations", gin.WrapF(handler.AIImagesGenerations))
 	v1.POST("/images/edits", gin.WrapF(handler.AIImagesEdits))
 	v1.POST("/chat/completions", gin.WrapF(handler.AIChatCompletions))
 	v1.POST("/videos", gin.WrapF(handler.AIVideos))
+	v1.POST("/media/references", gin.WrapF(handler.UploadReferenceMedia))
 	v1.GET("/videos/:id", func(c *gin.Context) {
 		handler.AIVideo(c.Writer, c.Request, c.Param("id"))
 	})
diff --git a/service/settings.go b/service/settings.go
index bbf088f..23b2913 100644
--- a/service/settings.go
+++ b/service/settings.go
@@ -180,7 +180,7 @@ func SelectModelChannel(modelName string) (model.ModelChannel, error) {
 
 func BuildModelChannelURL(channel model.ModelChannel, path string) string {
 	baseURL := strings.TrimRight(channel.BaseURL, "/")
-	if !strings.HasSuffix(baseURL, "/v1") {
+	if !strings.HasSuffix(baseURL, "/v1") && !strings.HasSuffix(baseURL, "/api/v3") && !strings.HasSuffix(baseURL, "/api/plan/v3") {
 		baseURL += "/v1"
 	}
 	return baseURL + path
diff --git a/web/src/services/api/image.ts b/web/src/services/api/image.ts
index 938fb85..dea317e 100644
--- a/web/src/services/api/image.ts
+++ b/web/src/services/api/image.ts
@@ -100,11 +100,17 @@ function parseImagePayload(payload: ImageApiResponse) {
 function readAxiosError(error: unknown, fallback: string) {
     if (axios.isAxiosError<{ error?: { message?: string }; msg?: string; code?: number }>(error)) {
         const responseData = error.response?.data;
-        return responseData?.msg || responseData?.error?.message || (error.response?.status ? `${fallback}：${error.response.status}` : fallback);
+        return responseData?.msg || responseData?.error?.message || readStatusError(error.response?.status, fallback);
     }
     return error instanceof Error ? error.message : fallback;
 }
 
+function readStatusError(status: number | undefined, fallback: string) {
+    if (status === 401 || status === 403) return "鉴权失败，请检查 API Key、套餐权限或模型权限";
+    if (status === 429) return "请求被限流或额度不足，请稍后重试";
+    return status ? `${fallback}：${status}` : fallback;
+}
+
 function parseStreamChunk(chunk: string, onDelta: (value: string) => void) {
     let deltaText = "";
     for (const eventBlock of chunk.split("\n\n")) {
diff --git a/web/src/services/api/video.ts b/web/src/services/api/video.ts
index 863bcab..56ed12a 100644
--- a/web/src/services/api/video.ts
+++ b/web/src/services/api/video.ts
@@ -1,29 +1,66 @@
 import axios from "axios";
 
 import { dataUrlToFile } from "@/lib/image-utils";
+import { getMediaBlob, uploadMediaFile, type UploadedFile } from "@/services/file-storage";
 import { imageToDataUrl } from "@/services/image-storage";
 import { buildApiUrl, type AiConfig } from "@/stores/use-config-store";
 import { useUserStore } from "@/stores/use-user-store";
 import type { ReferenceImage } from "@/types/image";
+import type { ReferenceVideo } from "@/types/media";
 
 type VideoResponse = { id: string; status?: string; error?: { message?: string } };
 type ApiVideoResponse = VideoResponse | { code?: number; data?: VideoResponse | null; msg?: string };
+type SeedanceTask = {
+    id: string;
+    status?: "queued" | "running" | "succeeded" | "failed" | "cancelled" | "expired";
+    error?: { code?: string; message?: string } | null;
+    content?: { video_url?: string; last_frame_url?: string } | null;
+};
+type ApiEnvelope<T> = T | { code?: number; data?: T | null; msg?: string };
+type ReferenceMediaUploadResponse = { id: string; url: string; mimeType: string; bytes: number };
+
+export type VideoGenerationResult = { blob?: Blob; url?: string; mimeType?: string };
 
 function aiApiUrl(config: AiConfig, path: string) {
     return config.channelMode === "remote" ? `/api/v1${path}` : buildApiUrl(config.baseUrl, path);
 }
 
-function aiHeaders(config: AiConfig) {
+function aiHeaders(config: AiConfig, contentType?: string) {
     const token = useUserStore.getState().token;
-    return config.channelMode === "remote" ? (token ? { Authorization: `Bearer ${token}` } : undefined) : { Authorization: `Bearer ${config.apiKey}` };
+    return config.channelMode === "remote"
+        ? {
+              ...(token ? { Authorization: `Bearer ${token}` } : {}),
+              ...(contentType ? { "Content-Type": contentType } : {}),
+          }
+        : {
+              Authorization: `Bearer ${config.apiKey}`,
+              ...(contentType ? { "Content-Type": contentType } : {}),
+          };
 }
 
 function refreshRemoteUser(config: AiConfig) {
     if (config.channelMode === "remote") void useUserStore.getState().hydrateUser();
 }
 
-export async function requestVideoGeneration(config: AiConfig, prompt: string, references: ReferenceImage[] = []) {
-    const model = config.model || config.videoModel;
+export async function requestVideoGeneration(config: AiConfig, prompt: string, references: ReferenceImage[] = [], videoReferences: ReferenceVideo[] = []): Promise<VideoGenerationResult> {
+    const model = (config.model || config.videoModel).trim();
+    assertVideoConfig(config, model);
+    if (isSeedanceConfig(config, model)) {
+        return requestSeedanceGeneration(config, model, prompt, references, videoReferences);
+    }
+    if (videoReferences.length) {
+        throw new Error("当前视频接口不支持参考视频，请切换到 Seedance 2.0 / 火山 Agent Plan 模型，或移除参考视频");
+    }
+    return requestOpenAIVideoGeneration(config, model, prompt, references);
+}
+
+export async function storeGeneratedVideo(result: VideoGenerationResult): Promise<UploadedFile> {
+    if (result.blob) return uploadMediaFile(result.blob, "video");
+    if (result.url) return { url: result.url, storageKey: "", bytes: 0, mimeType: result.mimeType || "video/mp4" };
+    throw new Error("视频接口没有返回可播放的视频");
+}
+
+async function requestOpenAIVideoGeneration(config: AiConfig, model: string, prompt: string, references: ReferenceImage[]) {
     const body = new FormData();
     body.append("model", model);
     body.append("prompt", prompt);
@@ -36,26 +73,136 @@ export async function requestVideoGeneration(config: AiConfig, prompt: string, r
     try {
         const created = unwrapVideoResponse((await axios.post<ApiVideoResponse>(aiApiUrl(config, "/videos"), body, { headers: aiHeaders(config) })).data);
         if (!created.id) throw new Error("视频接口没有返回任务 ID");
-        for (;;) {
+        for (let attempt = 0; attempt < 120; attempt += 1) {
             const video = unwrapVideoResponse((await axios.get<ApiVideoResponse>(aiApiUrl(config, `/videos/${created.id}`), { headers: aiHeaders(config), params: config.channelMode === "remote" ? { model } : undefined })).data);
             if (video.status === "completed") break;
             if (video.status === "failed" || video.status === "cancelled") throw new Error(video.error?.message || "视频生成失败");
-            await new Promise((resolve) => setTimeout(resolve, 2500));
+            if (attempt === 119) throw new Error("视频生成超时，请稍后重试");
+            await delay(2500);
         }
         const content = await axios.get<Blob>(aiApiUrl(config, `/videos/${created.id}/content`), { headers: aiHeaders(config), params: config.channelMode === "remote" ? { model } : undefined, responseType: "blob" });
         await assertVideoBlob(content.data);
         refreshRemoteUser(config);
-        return content.data;
+        return { blob: content.data };
     } catch (error) {
         throw new Error(readAxiosError(error, "视频生成失败"));
     }
 }
 
+async function requestSeedanceGeneration(config: AiConfig, model: string, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[]) {
+    const content = await buildSeedanceContent(config, prompt, references, videoReferences);
+    if (!content.length) throw new Error("请输入视频提示词，或连接参考图片/视频");
+    const payload = {
+        model,
+        content,
+        ratio: normalizeSeedanceRatio(config.size),
+        resolution: normalizeVideoResolution(config.vquality),
+        duration: normalizeSeedanceDuration(config.videoSeconds),
+        watermark: false,
+    };
+
+    try {
+        const created = unwrapSeedanceTask((await axios.post<ApiEnvelope<SeedanceTask>>(seedanceApiUrl(config), payload, { headers: aiHeaders(config, "application/json") })).data);
+        if (!created.id) throw new Error("Seedance 接口没有返回任务 ID");
+        for (let attempt = 0; attempt < 120; attempt += 1) {
+            const task = unwrapSeedanceTask((await axios.get<ApiEnvelope<SeedanceTask>>(seedanceApiUrl(config, created.id), { headers: aiHeaders(config), params: config.channelMode === "remote" ? { model } : undefined })).data);
+            if (task.status === "succeeded") {
+                const url = task.content?.video_url;
+                if (!url) throw new Error("Seedance 任务成功但没有返回视频 URL");
+                refreshRemoteUser(config);
+                return videoResultFromUrl(url);
+            }
+            if (task.status === "failed" || task.status === "cancelled" || task.status === "expired") throw new Error(task.error?.message || `Seedance 视频生成${task.status === "expired" ? "超时" : "失败"}`);
+            if (attempt === 119) throw new Error("Seedance 视频生成超时，请稍后重试");
+            await delay(5000);
+        }
+        throw new Error("Seedance 视频生成超时，请稍后重试");
+    } catch (error) {
+        throw new Error(readAxiosError(error, "Seedance 视频生成失败"));
+    }
+}
+
+function seedanceApiUrl(config: AiConfig, taskId?: string) {
+    if (config.channelMode === "remote") return taskId ? `/api/v1/videos/${encodeURIComponent(taskId)}` : "/api/v1/videos";
+    return buildApiUrl(config.baseUrl, `/contents/generations/tasks${taskId ? `/${encodeURIComponent(taskId)}` : ""}`);
+}
+
+async function buildSeedanceContent(config: AiConfig, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[]) {
+    const content: Array<Record<string, unknown>> = [];
+    const text = prompt.trim();
+    if (text) content.push({ type: "text", text });
+    for (const image of references.slice(0, 7)) {
+        content.push({ type: "image_url", image_url: { url: await resolveSeedanceImageUrl(config, image) }, role: "reference_image" });
+    }
+    for (const video of videoReferences.slice(0, 3)) {
+        content.push({ type: "video_url", video_url: { url: await resolveSeedanceVideoUrl(video) }, role: "reference_video" });
+    }
+    return content;
+}
+
+async function resolveSeedanceImageUrl(config: AiConfig, image: ReferenceImage) {
+    const directUrl = image.url || image.dataUrl;
+    if (isPublicMediaUrl(directUrl) || directUrl.startsWith("asset://")) return directUrl;
+    const dataUrl = await imageToDataUrl(image);
+    if (!dataUrl) throw new Error("参考图读取失败，请换一张图片或重新上传");
+    if (config.channelMode === "remote") {
+        return uploadReferenceMedia(dataUrlToFile({ ...image, dataUrl }));
+    }
+    return dataUrl;
+}
+
+async function resolveSeedanceVideoUrl(video: ReferenceVideo) {
+    if (isPublicMediaUrl(video.url) || video.url.startsWith("asset://")) return video.url;
+    let blob: Blob | null = null;
+    if (video.storageKey) blob = await getMediaBlob(video.storageKey);
+    if (!blob && video.url?.startsWith("blob:")) blob = await (await fetch(video.url)).blob();
+    if (!blob) throw new Error("参考视频必须是公网 URL、素材 ID，或本地已保存的视频");
+    const file = new File([blob], video.name || "reference-video.mp4", { type: video.type || blob.type || "video/mp4" });
+    return uploadReferenceMedia(file);
+}
+
+async function uploadReferenceMedia(file: File) {
+    const token = useUserStore.getState().token;
+    if (!token) throw new Error("使用本地参考素材需要先登录，并在服务端配置 PUBLIC_BASE_URL");
+    const body = new FormData();
+    body.append("file", file, file.name);
+    const response = await axios.post<ApiEnvelope<ReferenceMediaUploadResponse>>("/api/v1/media/references", body, { headers: { Authorization: `Bearer ${token}` } });
+    const payload = unwrapEnvelope(response.data, "参考素材上传失败");
+    if (!payload.url) throw new Error("参考素材上传后没有返回公网 URL");
+    return payload.url;
+}
+
+async function videoResultFromUrl(url: string): Promise<VideoGenerationResult> {
+    try {
+        const response = await axios.get<Blob>(url, { responseType: "blob" });
+        await assertVideoBlob(response.data);
+        return { blob: response.data };
+    } catch {
+        return { url, mimeType: "video/mp4" };
+    }
+}
+
+function isSeedanceConfig(config: AiConfig, model: string) {
+    const value = `${model} ${config.baseUrl}`.toLowerCase();
+    return value.includes("seedance") || value.includes("doubao-seedance") || value.includes("ark.cn-beijing.volces.com/api/plan/v3");
+}
+
+function assertVideoConfig(config: AiConfig, model: string) {
+    if (!model) throw new Error("请先配置视频模型");
+    if (config.channelMode === "local" && !config.baseUrl.trim()) throw new Error("请先配置 Base URL");
+    if (config.channelMode === "local" && !config.apiKey.trim()) throw new Error("请先配置 API Key");
+}
+
 function normalizeVideoSeconds(value: string) {
     const seconds = Math.floor(Number(value) || 6);
     return String(Math.max(1, Math.min(20, seconds)));
 }
 
+function normalizeSeedanceDuration(value: string) {
+    const seconds = Math.floor(Number(value) || 5);
+    return Math.max(4, Math.min(15, seconds));
+}
+
 function normalizeVideoSize(value: string) {
     if (value === "auto") return null;
     const size = value || "1280x720";
@@ -63,6 +210,26 @@ function normalizeVideoSize(value: string) {
     return ["9:16", "2:3", "3:4"].includes(size) ? "720x1280" : "1280x720";
 }
 
+function normalizeSeedanceRatio(value: string) {
+    if (!value || value === "auto") return "adaptive";
+    if (["16:9", "4:3", "1:1", "3:4", "9:16", "21:9", "adaptive"].includes(value)) return value;
+    const match = value.match(/^(\d+)x(\d+)$/);
+    if (!match) return "adaptive";
+    const width = Number(match[1]);
+    const height = Number(match[2]);
+    if (!width || !height) return "adaptive";
+    const ratio = width / height;
+    const options = [
+        ["16:9", 16 / 9],
+        ["4:3", 4 / 3],
+        ["1:1", 1],
+        ["3:4", 3 / 4],
+        ["9:16", 9 / 16],
+        ["21:9", 21 / 9],
+    ] as const;
+    return options.reduce((best, item) => (Math.abs(item[1] - ratio) < Math.abs(best[1] - ratio) ? item : best), options[0])[0];
+}
+
 function normalizeVideoResolution(value: string) {
     if (value === "low") return "480p";
     if (value === "auto" || value === "high" || value === "medium") return "720p";
@@ -71,30 +238,53 @@ function normalizeVideoResolution(value: string) {
 }
 
 function unwrapVideoResponse(payload: ApiVideoResponse) {
-    if (!payload) throw new Error("接口没有返回视频任务");
-    if ("code" in payload && typeof payload.code === "number") {
+    return unwrapEnvelope(payload, "接口没有返回视频任务");
+}
+
+function unwrapSeedanceTask(payload: ApiEnvelope<SeedanceTask>) {
+    return unwrapEnvelope(payload, "Seedance 接口没有返回任务");
+}
+
+function unwrapEnvelope<T>(payload: ApiEnvelope<T>, emptyMessage: string): T {
+    if (!payload) throw new Error(emptyMessage);
+    if (typeof payload === "object" && "code" in payload && typeof payload.code === "number") {
         if (payload.code !== 0) throw new Error(payload.msg || "请求失败");
-        if (!payload.data) throw new Error("接口没有返回视频任务");
+        if (!payload.data) throw new Error(emptyMessage);
         return payload.data;
     }
-    return payload;
+    return payload as T;
 }
 
 function readAxiosError(error: unknown, fallback: string) {
     if (axios.isAxiosError<{ error?: { message?: string }; msg?: string; code?: number }>(error)) {
         const responseData = error.response?.data;
-        return responseData?.msg || responseData?.error?.message || (error.response?.status ? `${fallback}：${error.response.status}` : fallback);
+        return responseData?.msg || responseData?.error?.message || statusMessage(error.response?.status, fallback);
     }
     return error instanceof Error ? error.message : fallback;
 }
 
+function statusMessage(status: number | undefined, fallback: string) {
+    if (status === 401 || status === 403) return "鉴权失败，请检查 API Key、套餐权限或模型权限";
+    if (status === 429) return "请求被限流或额度不足，请稍后重试";
+    return status ? `${fallback}（${status}）` : fallback;
+}
+
 async function assertVideoBlob(blob: Blob) {
     if (!blob.type.includes("json")) return;
-    let payload: { code?: number; msg?: string };
+    let payload: { code?: number; msg?: string; error?: { message?: string } };
     try {
-        payload = JSON.parse(await blob.text()) as { code?: number; msg?: string };
+        payload = JSON.parse(await blob.text()) as { code?: number; msg?: string; error?: { message?: string } };
     } catch {
         return;
     }
     if (typeof payload.code === "number" && payload.code !== 0) throw new Error(payload.msg || "视频下载失败");
+    if (payload.error?.message) throw new Error(payload.error.message);
+}
+
+function isPublicMediaUrl(value: string) {
+    return /^https?:\/\//i.test(value || "");
+}
+
+function delay(ms: number) {
+    return new Promise((resolve) => setTimeout(resolve, ms));
 }
diff --git a/web/src/stores/use-config-store.ts b/web/src/stores/use-config-store.ts
index c4ba0f5..08f2912 100644
--- a/web/src/stores/use-config-store.ts
+++ b/web/src/stores/use-config-store.ts
@@ -126,6 +126,6 @@ export function useEffectiveConfig() {
 
 export function buildApiUrl(baseUrl: string, path: string) {
     const normalizedBaseUrl = baseUrl.trim().replace(/\/+$/, "");
-    const apiBaseUrl = normalizedBaseUrl.endsWith("/v1") ? normalizedBaseUrl : `${normalizedBaseUrl}/v1`;
+    const apiBaseUrl = normalizedBaseUrl.endsWith("/v1") || normalizedBaseUrl.endsWith("/api/v3") || normalizedBaseUrl.endsWith("/api/plan/v3") ? normalizedBaseUrl : `${normalizedBaseUrl}/v1`;
     return `${apiBaseUrl}${path}`;
 }
diff --git a/web/src/types/media.ts b/web/src/types/media.ts
new file mode 100644
index 0000000..6d95955
--- /dev/null
+++ b/web/src/types/media.ts
@@ -0,0 +1,7 @@
+export type ReferenceVideo = {
+    id: string;
+    name: string;
+    type: string;
+    url: string;
+    storageKey?: string;
+};

From 1b2d473d9cd91dcf359cb178e27a1289a9e0300b Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Tue, 26 May 2026 23:10:49 +0800
Subject: [PATCH 03/14] feat: add video node preview

---
 web/src/app/(user)/assets/asset-transfer.ts   |  3 +-
 .../(user)/canvas/[id]/canvas-client-page.tsx | 28 ++++++---
 .../components/canvas-config-node-panel.tsx   | 40 +++++++++++-
 .../components/canvas-node-generation.ts      | 20 ++++++
 .../(user)/canvas/components/canvas-node.tsx  |  2 +-
 web/src/app/(user)/image/page.tsx             |  4 +-
 web/src/app/(user)/video/page.tsx             | 62 ++++++++++++++++---
 web/src/components/model-picker.tsx           |  2 +-
 8 files changed, 137 insertions(+), 24 deletions(-)

diff --git a/web/src/app/(user)/assets/asset-transfer.ts b/web/src/app/(user)/assets/asset-transfer.ts
index dd50cd3..bb0a2b8 100644
--- a/web/src/app/(user)/assets/asset-transfer.ts
+++ b/web/src/app/(user)/assets/asset-transfer.ts
@@ -26,7 +26,8 @@ export async function exportAssets(assets: Asset[]) {
 
     await Promise.all(
         assets.map(async (asset) => {
-            const storageKey = asset.kind === "image" || asset.kind === "video" ? asset.data.storageKey : undefined;
+            if (asset.kind !== "image" && asset.kind !== "video") return;
+            const storageKey = asset.data.storageKey;
             if (!storageKey) return;
             const blob = asset.kind === "image" ? await getImageBlob(storageKey) : await getMediaBlob(storageKey);
             if (!blob) return;
diff --git a/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx b/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx
index 0f5671b..20020e0 100644
--- a/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx
+++ b/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx
@@ -7,7 +7,7 @@ import { Home, ImageIcon, Images, List, Menu, MessageSquare, Plus, Redo2, Settin
 import { saveAs } from "file-saver";
 
 import { requestEdit, requestGeneration, requestImageQuestion } from "@/services/api/image";
-import { requestVideoGeneration } from "@/services/api/video";
+import { requestVideoGeneration, storeGeneratedVideo } from "@/services/api/video";
 import { defaultConfig, type AiConfig, useConfigStore, useEffectiveConfig } from "@/stores/use-config-store";
 import { resolveImageUrl, uploadImage, type UploadedImage } from "@/services/image-storage";
 import { resolveMediaUrl, uploadMediaFile, type UploadedFile } from "@/services/file-storage";
@@ -1749,14 +1749,14 @@ function InfiniteCanvasPage() {
                         position: isEmptyVideoNode ? sourceNode.position : { x: parent.x + (sourceNode?.width || spec.width) + 96, y: parent.y },
                         width: isEmptyVideoNode ? sourceNode.width : spec.width,
                         height: isEmptyVideoNode ? sourceNode.height : spec.height,
-                        metadata: { prompt: effectivePrompt, status: NODE_STATUS_LOADING, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, references: generationContext.referenceImages.map(referenceUrl).filter((url): url is string => Boolean(url)) },
+                        metadata: { prompt: effectivePrompt, status: NODE_STATUS_LOADING, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, references: generationReferenceUrls(generationContext) },
                     };
                     pendingChildIds = [videoId];
                     setNodes((prev) => (isEmptyVideoNode ? prev.map((node) => (node.id === nodeId ? { ...node, ...videoNode } : node)) : [...prev.map((node) => (node.id === nodeId ? { ...node, metadata: { ...node.metadata, status: NODE_STATUS_SUCCESS } } : node)), videoNode]));
                     if (!isEmptyVideoNode) setConnections((prev) => [...prev, { id: nanoid(), fromNodeId: nodeId, toNodeId: videoId }]);
-                    const video = await uploadMediaFile(await requestVideoGeneration(generationConfig, effectivePrompt, generationContext.referenceImages), "video");
+                    const video = await storeGeneratedVideo(await requestVideoGeneration(generationConfig, effectivePrompt, generationContext.referenceImages, generationContext.referenceVideos));
                     const videoSize = fitNodeSize(video.width || spec.width, video.height || spec.height, VIDEO_NODE_MAX_WIDTH, VIDEO_NODE_MAX_HEIGHT);
-                    setNodes((prev) => prev.map((node) => (node.id === videoId ? { ...node, width: videoSize.width, height: videoSize.height, position: { x: node.position.x + node.width / 2 - videoSize.width / 2, y: node.position.y + node.height / 2 - videoSize.height / 2 }, metadata: { ...node.metadata, ...videoMetadata(video), prompt: effectivePrompt, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, references: generationContext.referenceImages.map(referenceUrl).filter((url): url is string => Boolean(url)) } } : node)));
+                    setNodes((prev) => prev.map((node) => (node.id === videoId ? { ...node, width: videoSize.width, height: videoSize.height, position: { x: node.position.x + node.width / 2 - videoSize.width / 2, y: node.position.y + node.height / 2 - videoSize.height / 2 }, metadata: { ...node.metadata, ...videoMetadata(video), prompt: effectivePrompt, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, references: generationReferenceUrls(generationContext) } } : node)));
                     return;
                 }
 
@@ -1857,6 +1857,7 @@ function InfiniteCanvasPage() {
                 setNodes((prev) => prev.map((item) => (item.id === node.id ? { ...item, metadata: { ...item.metadata, status: NODE_STATUS_ERROR, errorDetails: "参考图片已丢失，无法继续重试" } } : item)));
                 return;
             }
+            const retryImages = retryReferenceImages || [];
 
             setRunningNodeId(node.id);
             setNodes((prev) => prev.map((item) => (item.id === node.id ? { ...item, metadata: { ...item.metadata, status: NODE_STATUS_LOADING, errorDetails: undefined } } : item)));
@@ -1873,19 +1874,19 @@ function InfiniteCanvasPage() {
                     return;
                 }
                 if (node.type === CanvasNodeType.Video) {
-                    const video = await uploadMediaFile(await requestVideoGeneration(generationConfig, prompt, retryReferenceImages || []), "video");
+                    const video = await storeGeneratedVideo(await requestVideoGeneration(generationConfig, prompt, retryImages, context?.referenceVideos || []));
                     const videoSize = fitNodeSize(video.width || node.width, video.height || node.height, VIDEO_NODE_MAX_WIDTH, VIDEO_NODE_MAX_HEIGHT);
                     setNodes((prev) => prev.map((item) => (item.id === node.id ? { ...item, width: videoSize.width, height: videoSize.height, position: { x: item.position.x + item.width / 2 - videoSize.width / 2, y: item.position.y + item.height / 2 - videoSize.height / 2 }, metadata: { ...item.metadata, ...videoMetadata(video), prompt, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality } } : item)));
                     return;
                 }
 
-                const image = useReferenceImages ? await requestEdit(generationConfig, prompt, retryReferenceImages).then((items) => items[0]) : await requestGeneration(generationConfig, prompt).then((items) => items[0]);
+                const image = useReferenceImages ? await requestEdit(generationConfig, prompt, retryImages).then((items) => items[0]) : await requestGeneration(generationConfig, prompt).then((items) => items[0]);
                 const uploadedImage = await uploadImage(image.dataUrl);
                 const imageConfig = NODE_DEFAULT_SIZE[CanvasNodeType.Image];
                 const imageSize = fitNodeSize(uploadedImage.width, uploadedImage.height, imageConfig.width, imageConfig.height);
                 const generationMetadata = savedImageMetadata?.generationType
                     ? { generationType: savedImageMetadata.generationType, model: generationConfig.model, size: generationConfig.size, quality: generationConfig.quality, count: savedImageMetadata.count || 1, references: savedImageMetadata.references }
-                    : buildImageGenerationMetadata(useReferenceImages ? "edit" : "generation", generationConfig, 1, retryReferenceImages || []);
+                    : buildImageGenerationMetadata(useReferenceImages ? "edit" : "generation", generationConfig, 1, retryImages);
                 setNodes((prev) =>
                     prev.map((item) =>
                         item.id === node.id
@@ -2521,6 +2522,13 @@ function referenceUrl(image: ReferenceImage) {
     return image.storageKey || image.url || (!image.dataUrl.startsWith("data:") ? image.dataUrl : undefined);
 }
 
+function generationReferenceUrls(context: { referenceImages: ReferenceImage[]; referenceVideos: Array<{ storageKey?: string; url?: string }> }) {
+    return [
+        ...context.referenceImages.map(referenceUrl).filter((url): url is string => Boolean(url)),
+        ...context.referenceVideos.map((video) => video.storageKey || video.url).filter((url): url is string => Boolean(url)),
+    ];
+}
+
 async function resolveMetadataReferences(metadata: CanvasNodeMetadata) {
     if (metadata.generationType !== "edit") return [];
     if (!metadata.references?.length) return null;
@@ -2574,9 +2582,10 @@ function getGenerationCount(count: string) {
 }
 
 function applyNodeConfigPatch(node: CanvasNodeData, patch: Partial<CanvasNodeData["metadata"]>) {
-    const next = { ...node, metadata: { ...node.metadata, ...(patch || {}) } };
+    const safePatch = patch || {};
+    const next = { ...node, metadata: { ...node.metadata, ...safePatch } };
     const spec = node.type === CanvasNodeType.Video ? NODE_DEFAULT_SIZE[CanvasNodeType.Video] : NODE_DEFAULT_SIZE[CanvasNodeType.Image];
-    const size = typeof patch.size === "string" && !node.metadata?.content ? nodeSizeFromRatio(patch.size, spec.width, spec.height) : null;
+    const size = typeof safePatch.size === "string" && !node.metadata?.content ? nodeSizeFromRatio(safePatch.size, spec.width, spec.height) : null;
     return size && (node.type === CanvasNodeType.Image || node.type === CanvasNodeType.Video) ? { ...next, ...size, position: { x: node.position.x + node.width / 2 - size.width / 2, y: node.position.y + node.height / 2 - size.height / 2 } } : next;
 }
 
@@ -2595,6 +2604,7 @@ function getInputSummary(inputs: NodeGenerationInput[]) {
     return {
         textCount: inputs.filter((input) => input.type === "text").length,
         imageCount: inputs.filter((input) => input.type === "image").length,
+        videoCount: inputs.filter((input) => input.type === "video").length,
     };
 }
 
diff --git a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
index 8f5c117..216b814 100644
--- a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
@@ -18,7 +18,7 @@ import type { CanvasGenerationMode, CanvasNodeData, CanvasNodeMetadata } from ".
 type CanvasConfigNodePanelProps = {
     node: CanvasNodeData;
     isRunning: boolean;
-    inputSummary: { textCount: number; imageCount: number };
+    inputSummary: { textCount: number; imageCount: number; videoCount: number };
     inputs: NodeGenerationInput[];
     onConfigChange: (nodeId: string, patch: Partial<CanvasNodeMetadata>) => void;
     onTextInputChange: (nodeId: string, content: string) => void;
@@ -41,6 +41,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
     const chipStyle = { background: theme.node.fill, borderColor: theme.node.stroke, color: theme.node.text };
     const textInputs = inputs.filter((input) => input.type === "text");
     const imageInputs = inputs.filter((input) => input.type === "image");
+    const videoInputs = inputs.filter((input) => input.type === "video");
 
     const moveInput = (input: NodeGenerationInput, offset: number) => {
         const sameTypeInputs = inputs.filter((item) => item.type === input.type);
@@ -112,6 +113,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
             <div className="mb-2 flex flex-wrap gap-1.5" onMouseDown={(event) => event.stopPropagation()}>
                 <InputChip label="提示词" value={`${inputSummary.textCount} 个`} style={chipStyle} />
                 <InputChip label="参考图" value={`${inputSummary.imageCount} 张`} style={chipStyle} />
+                <InputChip label="参考视频" value={`${inputSummary.videoCount} 个`} style={chipStyle} />
                 <button type="button" className="inline-flex h-7 cursor-pointer items-center gap-1 rounded-md border px-2 text-[11px]" style={chipStyle} onClick={() => setPreviewOpen(true)}>
                     <Eye className="size-3.5" />
                     预览
@@ -130,7 +132,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
             <Button
                 type="primary"
                 className="mt-auto !h-9 !w-full !cursor-pointer !rounded-lg"
-                disabled={isRunning || (!inputSummary.textCount && !inputSummary.imageCount)}
+                disabled={isRunning || (!inputSummary.textCount && !inputSummary.imageCount && !inputSummary.videoCount)}
                 onMouseDown={(event) => event.stopPropagation()}
                 onClick={() => onGenerate(node.id)}
             >
@@ -171,6 +173,15 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
                                     </div>
                                 </PreviewSection>
                             </div>
+                            <div className="shrink-0">
+                                <PreviewSection title="参考视频" count={videoInputs.length} empty="暂无参考视频">
+                                    <div className="thin-scrollbar flex gap-1.5 overflow-x-auto pb-1">
+                                        {videoInputs.map((input, index) => (
+                                            <VideoSortCard key={input.nodeId} input={input} videoIndex={index} videoTotal={videoInputs.length} theme={theme} onMove={moveInput} />
+                                        ))}
+                                    </div>
+                                </PreviewSection>
+                            </div>
                             <div className="grid min-h-0 flex-1 grid-cols-2 gap-3 overflow-hidden">
                                 <div className="thin-scrollbar min-h-0 overflow-y-auto pr-1.5">
                                     <PreviewSection title="文本提示词" count={textInputs.length} empty="暂无文本提示词">
@@ -288,6 +299,31 @@ function ImageSortCard({
     );
 }
 
+function VideoSortCard({
+    input,
+    videoIndex,
+    videoTotal,
+    theme,
+    onMove,
+}: {
+    input: NodeGenerationInput;
+    videoIndex: number;
+    videoTotal: number;
+    theme: (typeof canvasThemes)[keyof typeof canvasThemes];
+    onMove: (input: NodeGenerationInput, offset: number) => void;
+}) {
+    if (!input.video) return null;
+    return (
+        <div className="w-32 shrink-0 overflow-hidden rounded-lg border" style={{ background: theme.node.fill, borderColor: theme.node.stroke }}>
+            <div className="relative">
+                <video src={input.video.url} className="aspect-video w-full bg-black object-cover" muted preload="metadata" />
+                <span className="absolute left-1 top-1 rounded bg-black/50 px-1 py-0.5 text-[9px] font-medium text-white">{videoIndex + 1}</span>
+                <HorizontalOrderButtons index={videoIndex} total={videoTotal} onMove={(offset) => onMove(input, offset)} />
+            </div>
+        </div>
+    );
+}
+
 function VerticalOrderButtons({ index, total, onMove }: { index: number; total: number; onMove: (offset: number) => void }) {
     return (
         <>
diff --git a/web/src/app/(user)/canvas/components/canvas-node-generation.ts b/web/src/app/(user)/canvas/components/canvas-node-generation.ts
index 83513f2..3b4c09a 100644
--- a/web/src/app/(user)/canvas/components/canvas-node-generation.ts
+++ b/web/src/app/(user)/canvas/components/canvas-node-generation.ts
@@ -1,12 +1,15 @@
 import type { ChatCompletionMessage } from "@/services/api/image";
 import type { ReferenceImage } from "@/types/image";
+import type { ReferenceVideo } from "@/types/media";
 import { CanvasNodeType, type CanvasConnection, type CanvasNodeData } from "../types";
 
 export type NodeGenerationContext = {
     prompt: string;
     referenceImages: ReferenceImage[];
+    referenceVideos: ReferenceVideo[];
     textCount: number;
     imageCount: number;
+    videoCount: number;
 };
 
 export type NodeGenerationInput = {
@@ -15,6 +18,7 @@ export type NodeGenerationInput = {
     title: string;
     text?: string;
     image?: ReferenceImage;
+    video?: ReferenceVideo;
 };
 
 export function buildNodeGenerationContext(nodeId: string, nodes: CanvasNodeData[], connections: CanvasConnection[], prompt: string): NodeGenerationContext {
@@ -24,12 +28,15 @@ export function buildNodeGenerationContext(nodeId: string, nodes: CanvasNodeData
         .filter(Boolean)
         .join("\n\n");
     const referenceImages = inputs.map((input) => input.image).filter((image): image is ReferenceImage => Boolean(image));
+    const referenceVideos = inputs.map((input) => input.video).filter((video): video is ReferenceVideo => Boolean(video));
 
     return {
         prompt: upstreamText ? `${prompt}\n\n${upstreamText}` : prompt,
         referenceImages,
+        referenceVideos,
         textCount: inputs.filter((input) => input.type === "text").length,
         imageCount: referenceImages.length,
+        videoCount: referenceVideos.length,
     };
 }
 
@@ -37,6 +44,8 @@ export function buildNodeGenerationInputs(nodeId: string, nodes: CanvasNodeData[
     return getOrderedUpstreamNodes(nodeId, nodes, connections).flatMap((node): NodeGenerationInput[] => {
         const image = readReferenceImage(node);
         if (image) return [{ nodeId: node.id, type: "image" as const, title: node.title, image }];
+        const video = readReferenceVideo(node);
+        if (video) return [{ nodeId: node.id, type: "video" as const, title: node.title, video }];
         const text = readNodeTextInput(node);
         if (text) return [{ nodeId: node.id, type: "text" as const, title: node.title, text }];
         return [];
@@ -77,6 +86,17 @@ function readReferenceImage(node: CanvasNodeData): ReferenceImage | null {
     };
 }
 
+function readReferenceVideo(node: CanvasNodeData): ReferenceVideo | null {
+    if (node.type !== CanvasNodeType.Video || !node.metadata?.content) return null;
+    return {
+        id: node.id,
+        name: `${node.title || node.id}.mp4`,
+        type: node.metadata.mimeType || "video/mp4",
+        url: node.metadata.content,
+        storageKey: node.metadata.storageKey,
+    };
+}
+
 function getOrderedUpstreamNodes(nodeId: string, nodes: CanvasNodeData[], connections: CanvasConnection[]) {
     const target = nodes.find((node) => node.id === nodeId);
     const upstreamNodes = connections
diff --git a/web/src/app/(user)/canvas/components/canvas-node.tsx b/web/src/app/(user)/canvas/components/canvas-node.tsx
index 6f63b9d..7074825 100644
--- a/web/src/app/(user)/canvas/components/canvas-node.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-node.tsx
@@ -466,7 +466,7 @@ function VideoNodeContent({ node, theme }: NodeContentRendererProps) {
         return (
             <div className="flex h-full w-full flex-col items-center justify-center gap-3" style={{ color: theme.node.placeholder }}>
                 <Video className="size-7 opacity-35" />
-                <span className="text-sm">视频节点(未开发完，请勿使用)</span>
+                <span className="text-sm">空视频节点</span>
             </div>
         );
     return <video src={node.metadata.content} controls className="h-full w-full rounded-[18px] bg-black object-contain" data-canvas-no-zoom />;
diff --git a/web/src/app/(user)/image/page.tsx b/web/src/app/(user)/image/page.tsx
index 073898f..34be005 100644
--- a/web/src/app/(user)/image/page.tsx
+++ b/web/src/app/(user)/image/page.tsx
@@ -218,9 +218,11 @@ export default function ImagePage() {
     const insertPickedAsset = async (payload: InsertAssetPayload) => {
         if (payload.kind === "text") {
             setPrompt(payload.content);
-        } else {
+        } else if (payload.kind === "image") {
             const stored = await uploadImage(payload.dataUrl);
             setReferences((value) => [...value, { id: nanoid(), name: payload.title, type: stored.mimeType, dataUrl: stored.url, storageKey: stored.storageKey }]);
+        } else {
+            message.warning("生图工作台只能使用文本或图片素材");
         }
         setAssetPickerOpen(false);
     };
diff --git a/web/src/app/(user)/video/page.tsx b/web/src/app/(user)/video/page.tsx
index 85f9c42..9205d70 100644
--- a/web/src/app/(user)/video/page.tsx
+++ b/web/src/app/(user)/video/page.tsx
@@ -15,11 +15,12 @@ import { canvasThemes } from "@/lib/canvas-theme";
 import { formatBytes, formatDuration } from "@/lib/image-utils";
 import { deleteStoredMedia, resolveMediaUrl, uploadMediaFile } from "@/services/file-storage";
 import { resolveImageUrl, uploadImage } from "@/services/image-storage";
-import { requestVideoGeneration } from "@/services/api/video";
+import { requestVideoGeneration, storeGeneratedVideo } from "@/services/api/video";
 import { useAssetStore } from "@/stores/use-asset-store";
 import { useConfigStore, useEffectiveConfig, type AiConfig } from "@/stores/use-config-store";
 import { useThemeStore } from "@/stores/use-theme-store";
 import type { ReferenceImage } from "@/types/image";
+import type { ReferenceVideo } from "@/types/media";
 
 type GeneratedVideo = {
     id: string;
@@ -48,6 +49,7 @@ type GenerationLog = {
     model: string;
     config: GenerationLogConfig;
     references: ReferenceImage[];
+    videoReferences: ReferenceVideo[];
     durationMs: number;
     size: string;
     resolution: string;
@@ -75,6 +77,7 @@ export default function VideoPage() {
     const addAsset = useAssetStore((state) => state.addAsset);
     const [prompt, setPrompt] = useState("");
     const [references, setReferences] = useState<ReferenceImage[]>([]);
+    const [videoReferences, setVideoReferences] = useState<ReferenceVideo[]>([]);
     const [results, setResults] = useState<GenerationResult[]>([]);
     const [logs, setLogs] = useState<GenerationLog[]>([]);
     const [running, setRunning] = useState(false);
@@ -102,14 +105,23 @@ export default function VideoPage() {
     }, []);
 
     const addReferences = async (files?: FileList | null) => {
-        const imageFiles = Array.from(files || []).filter((file) => file.type.startsWith("image/")).slice(0, 7 - references.length);
+        const selectedFiles = Array.from(files || []);
+        const imageFiles = selectedFiles.filter((file) => file.type.startsWith("image/")).slice(0, 7 - references.length);
+        const videoFiles = selectedFiles.filter((file) => file.type.startsWith("video/")).slice(0, 3 - videoReferences.length);
         const nextReferences = await Promise.all(
             imageFiles.map(async (file) => {
                 const image = await uploadImage(file);
                 return { id: nanoid(), name: file.name, type: image.mimeType, dataUrl: image.url, storageKey: image.storageKey };
             }),
         );
+        const nextVideoReferences = await Promise.all(
+            videoFiles.map(async (file) => {
+                const video = await uploadMediaFile(file, "video-reference");
+                return { id: nanoid(), name: file.name, type: video.mimeType, url: video.url, storageKey: video.storageKey };
+            }),
+        );
         setReferences((value) => [...value, ...nextReferences].slice(0, 7));
+        setVideoReferences((value) => [...value, ...nextVideoReferences].slice(0, 3));
     };
 
     const addReferencesFromClipboard = async () => {
@@ -143,8 +155,7 @@ export default function VideoPage() {
         const batchStartedAt = performance.now();
         setStartedAt(batchStartedAt);
         try {
-            const blob = await requestVideoGeneration(snapshot.config, snapshot.text, snapshot.references);
-            const stored = await uploadMediaFile(blob, "video");
+            const stored = await storeGeneratedVideo(await requestVideoGeneration(snapshot.config, snapshot.text, snapshot.references, snapshot.videoReferences));
             const nextVideo: GeneratedVideo = {
                 id: nanoid(),
                 url: stored.url,
@@ -156,12 +167,12 @@ export default function VideoPage() {
                 mimeType: stored.mimeType,
             };
             setResults([{ id: nextVideo.id, status: "success", video: nextVideo }]);
-            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, durationMs: nextVideo.durationMs, status: "成功", video: nextVideo }));
+            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, videoReferences: snapshot.videoReferences, durationMs: nextVideo.durationMs, status: "成功", video: nextVideo }));
             message.success("视频已生成");
         } catch (error) {
             const errorMessage = error instanceof Error ? error.message : "生成失败";
             setResults([{ id: nanoid(), status: "failed", error: errorMessage }]);
-            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, durationMs: performance.now() - batchStartedAt, status: "失败", error: errorMessage }));
+            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, videoReferences: snapshot.videoReferences, durationMs: performance.now() - batchStartedAt, status: "失败", error: errorMessage }));
             message.error(errorMessage);
         } finally {
             setRunning(false);
@@ -179,7 +190,7 @@ export default function VideoPage() {
             openConfigDialog(true);
             return null;
         }
-        return { text, config: buildVideoConfig(effectiveConfig, model), references: [...references] };
+        return { text, config: buildVideoConfig(effectiveConfig, model), references: [...references], videoReferences: [...videoReferences] };
     };
 
     const retryResult = () => {
@@ -209,6 +220,8 @@ export default function VideoPage() {
         } else if (payload.kind === "image") {
             const stored = await uploadImage(payload.dataUrl);
             setReferences((value) => [...value, { id: nanoid(), name: payload.title, type: stored.mimeType, dataUrl: stored.url, storageKey: stored.storageKey }].slice(0, 7));
+        } else if (payload.kind === "video") {
+            setVideoReferences((value) => [...value, { id: nanoid(), name: payload.title, type: "video/mp4", url: payload.url, storageKey: payload.storageKey }].slice(0, 3));
         }
         setAssetPickerOpen(false);
     };
@@ -216,6 +229,7 @@ export default function VideoPage() {
     const createSession = () => {
         setPrompt("");
         setReferences([]);
+        setVideoReferences([]);
         setResults([]);
         setElapsedMs(0);
         setStartedAt(0);
@@ -248,6 +262,7 @@ export default function VideoPage() {
         setLogsOpen(false);
         setPrompt(log.prompt);
         setReferences(log.references || []);
+        setVideoReferences(log.videoReferences || []);
         if (log.config.videoModel || log.model) updateConfig("videoModel", log.config.videoModel || log.model);
         if (log.config.size) updateConfig("size", log.config.size);
         if (log.config.vquality) updateConfig("vquality", log.config.vquality);
@@ -317,6 +332,26 @@ export default function VideoPage() {
                                 </div>
                             </div>
 
+                            <div className="min-w-0">
+                                <div className="mb-2 flex items-center justify-between gap-3">
+                                    <span className="text-base font-semibold">参考视频</span>
+                                    <Button size="small" icon={<Upload className="size-3.5" />} onClick={() => fileInputRef.current?.click()}>
+                                        上传
+                                    </Button>
+                                </div>
+                                <div className="hover-scrollbar hover-scrollbar-hint flex min-h-24 w-full min-w-0 max-w-full gap-2 overflow-x-scroll overflow-y-hidden rounded-lg border border-dashed border-stone-300 p-2 pb-3 overscroll-x-contain dark:border-stone-700">
+                                    {videoReferences.map((item) => (
+                                        <div key={item.id} className="group relative h-20 w-32 shrink-0 overflow-hidden rounded-md border border-stone-200 bg-black dark:border-stone-800">
+                                            <video src={item.url} className="size-full object-cover" muted preload="metadata" />
+                                            <button type="button" className="absolute right-1 top-1 hidden size-6 items-center justify-center rounded bg-black/60 text-white group-hover:flex" onClick={() => setVideoReferences((value) => value.filter((ref) => ref.id !== item.id))} aria-label="移除参考视频">
+                                                <Trash2 className="size-3.5" />
+                                            </button>
+                                        </div>
+                                    ))}
+                                    {!videoReferences.length ? <div className="flex min-w-full items-center justify-center text-sm text-stone-500">暂无参考视频，最多 3 个</div> : null}
+                                </div>
+                            </div>
+
                             <div className="flex items-center justify-between rounded-lg border border-stone-200 bg-stone-50 px-3 py-2 text-sm dark:border-stone-800 dark:bg-stone-900 sm:hidden">
                                 <span className="truncate text-stone-500 dark:text-stone-400">
                                     {model} · {normalizeResolution(effectiveConfig.vquality)}p · {videoSizeLabel(effectiveConfig.size)} · {normalizeVideoSeconds(effectiveConfig.videoSeconds)}s
@@ -359,7 +394,7 @@ export default function VideoPage() {
             <input
                 ref={fileInputRef}
                 type="file"
-                accept="image/*"
+                accept="image/*,video/mp4,video/quicktime"
                 multiple
                 className="hidden"
                 onChange={(event) => {
@@ -542,6 +577,12 @@ async function readStoredLogs() {
 
 async function normalizeLog(log: Partial<GenerationLog>): Promise<GenerationLog> {
     const video = log.video?.storageKey ? { ...log.video, url: await resolveMediaUrl(log.video.storageKey, log.video.url) } : log.video;
+    const videoReferences = await Promise.all(
+        (log.videoReferences || []).map(async (item) => ({
+            ...item,
+            url: item.storageKey ? await resolveMediaUrl(item.storageKey, item.url) : item.url,
+        })),
+    );
     const references = await Promise.all(
         (log.references || []).map(async (item) => ({
             ...item,
@@ -558,6 +599,7 @@ async function normalizeLog(log: Partial<GenerationLog>): Promise<GenerationLog>
         model: log.model || config.videoModel || "",
         config,
         references,
+        videoReferences,
         durationMs: log.durationMs || 0,
         size: log.size || config.size || "",
         resolution: normalizeResolution(log.resolution || config.vquality || ""),
@@ -572,6 +614,7 @@ function serializeLog(log: GenerationLog): GenerationLog {
     return {
         ...log,
         references: log.references.map((item) => ({ ...item, dataUrl: item.storageKey ? "" : item.dataUrl })),
+        videoReferences: log.videoReferences.map((item) => (item.storageKey ? { ...item, url: "" } : item)),
         video: log.video?.storageKey ? { ...log.video, url: "" } : log.video,
     };
 }
@@ -586,7 +629,7 @@ function normalizeLogConfig(log: Partial<GenerationLog>): GenerationLogConfig {
     };
 }
 
-function buildLog({ prompt, model, config, references, durationMs, status, video, error }: { prompt: string; model: string; config: AiConfig; references: ReferenceImage[]; durationMs: number; status: GenerationLog["status"]; video?: GeneratedVideo; error?: string }): GenerationLog {
+function buildLog({ prompt, model, config, references, videoReferences, durationMs, status, video, error }: { prompt: string; model: string; config: AiConfig; references: ReferenceImage[]; videoReferences: ReferenceVideo[]; durationMs: number; status: GenerationLog["status"]; video?: GeneratedVideo; error?: string }): GenerationLog {
     const logConfig = {
         model: config.model,
         videoModel: config.videoModel,
@@ -603,6 +646,7 @@ function buildLog({ prompt, model, config, references, durationMs, status, video
         model,
         config: logConfig,
         references,
+        videoReferences,
         durationMs,
         size: logConfig.size,
         resolution: logConfig.vquality,
diff --git a/web/src/components/model-picker.tsx b/web/src/components/model-picker.tsx
index 9e51980..8da2b3b 100644
--- a/web/src/components/model-picker.tsx
+++ b/web/src/components/model-picker.tsx
@@ -20,7 +20,7 @@ type ModelPickerProps = {
 export function ModelPicker({ config, value, onChange, className, fullWidth = false, placeholder = "选择模型", onMissingConfig }: ModelPickerProps) {
     const pickerId = useId();
     const [open, setOpen] = useState(false);
-    const options = useMemo(() => Array.from(new Set([...(config.channelMode === "local" ? [value] : []), ...config.models].filter(Boolean))), [config.channelMode, config.models, value]);
+    const options = useMemo(() => Array.from(new Set([...(config.channelMode === "local" ? [value] : []), ...config.models].filter((model): model is string => Boolean(model)))), [config.channelMode, config.models, value]);
     const current = value || "";
 
     useEffect(() => {

From e37597f4329f6d2790423e7e2d98cf74f806a2f0 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Tue, 26 May 2026 23:10:57 +0800
Subject: [PATCH 04/14] docs: add ark agentplan setup notes

---
 README.md                  |  2 +-
 docs/canvas-node-manual.md |  5 ++++-
 docs/features.md           | 22 ++++++++++++++++++++--
 docs/pending-test.md       |  3 +++
 4 files changed, 28 insertions(+), 4 deletions(-)

diff --git a/README.md b/README.md
index f7553b8..4557645 100644
--- a/README.md
+++ b/README.md
@@ -14,7 +14,7 @@
 ## 核心功能
 
 - 无限画布：多画布项目、节点拖拽缩放、连线、小地图、撤销重做、导入导出。
-- AI 创作：支持 OpenAI 兼容接口的文生图、图生图、参考图编辑和文本问答。
+- AI 创作：支持 OpenAI 兼容接口的文生图、图生图、参考图编辑、文本问答和视频生成；Seedance 2.0 可通过火山方舟 Agent Plan 接入。
 - 画布助手：围绕选中节点和上游节点对话、生图，并把结果插回画布。
 - 提示词库：抓取多个 GitHub 开源项目，按案例整理数百个图片提示词。
 
diff --git a/docs/canvas-node-manual.md b/docs/canvas-node-manual.md
index 845b9a6..48db9b8 100644
--- a/docs/canvas-node-manual.md
+++ b/docs/canvas-node-manual.md
@@ -45,7 +45,10 @@
 - 视频节点使用原生播放器展示内容，可在节点内直接播放、暂停和拖动进度。
 - 空视频节点下方对话框可输入提示词生成视频，结果会回填到当前节点。
 - 从文本、图片或配置节点创建视频生成时，会在右侧生成新的视频节点并自动连接。
-- 视频生成接口使用 OpenAI 兼容的 `POST /v1/videos`、`GET /v1/videos/{id}` 和 `GET /v1/videos/{id}/content`。
+- 生成配置节点的视频模式会读取上游文本作为 prompt，读取上游图片作为参考图，读取上游视频作为参考视频，并在输入预览里显示参考视频。
+- 视频生成接口支持 OpenAI 风格的 `POST /v1/videos`、`GET /v1/videos/{id}` 和 `GET /v1/videos/{id}/content`。
+- 使用火山方舟 Agent Plan / Seedance 2.0 时，Base URL 配置为 `https://ark.cn-beijing.volces.com/api/plan/v3`，模型名使用 Seedance 2.0 对应模型；系统会改用 `POST /contents/generations/tasks` 创建异步任务，并轮询 `GET /contents/generations/tasks/{id}`。
+- Seedance 参考视频必须是公网可访问 URL，或由本项目后端在配置 `PUBLIC_BASE_URL` 后上传并暴露的参考素材 URL。本地/内网地址无法被火山服务器拉取。
 
 ### 推荐流程
 
diff --git a/docs/features.md b/docs/features.md
index bec3f05..344dd90 100644
--- a/docs/features.md
+++ b/docs/features.md
@@ -52,13 +52,25 @@
 
 ## AI 生成
 
-项目通过前端直接请求 OpenAI 兼容接口：
+项目支持两种 AI 调用方式：
+
+- 本地直连：前端使用本地配置的 Base URL、API Key 和 Model 直接请求 OpenAI 兼容接口。
+- 后台渠道：前端请求本项目后端 `/api/v1/*` 代理接口，后端按模型选择管理后台配置的渠道。
+
+OpenAI 兼容图像和文本能力继续复用现有接口：
 
 - `/v1/images/generations`：文生图。
 - `/v1/images/edits`：图生图/参考图编辑。
 - `/v1/chat/completions`：文本问答和带图问答。
 - `/v1/models`：读取模型列表。
 
+视频能力支持两类接口：
+
+- OpenAI 风格视频：`POST /v1/videos`、`GET /v1/videos/{id}`、`GET /v1/videos/{id}/content`。
+- 火山方舟 Agent Plan / Seedance 2.0：Base URL 使用 `https://ark.cn-beijing.volces.com/api/plan/v3`，创建任务为 `POST /contents/generations/tasks`，查询任务为 `GET /contents/generations/tasks/{id}`，成功结果读取 `content.video_url`。
+
+Base URL 如果已经以 `/v1`、`/api/v3` 或 `/api/plan/v3` 结尾，系统不会再追加 `/v1`。因此 cpa 反代或火山方舟 Agent Plan 可以继续通过现有 Base URL + API Key + Model 方式配置，不需要新增火山生图 Provider。
+
 可配置项：
 
 - Base URL。
@@ -67,9 +79,13 @@
 - 图片质量。
 - 图片比例。
 - 生成数量。
+- 视频模型。
+- 视频比例、清晰度和时长。
 
 普通图片/文本节点可以直接输入提示词生成结果。生成配置节点可以读取上游节点内容，并按节点自己的配置批量生成多个图片或文本结果。生成配置节点支持预览当前提示词和参考图输入，并调整输入顺序。
 
+视频生成可从文本节点读取 prompt，从图片节点读取参考图，从视频节点读取参考视频。生成成功后会把视频插入画布为视频节点并使用原生播放器预览。Seedance 参考视频需要公网可访问 URL；本地上传视频会先通过 `/api/v1/media/references` 保存到服务端，再由 `PUBLIC_BASE_URL` 生成可供火山服务器拉取的公开链接。
+
 ## 画布助手
 
 画布右侧助手面板支持：
@@ -161,6 +177,8 @@
 ## 当前限制
 
 - 画布项目和“我的素材”目前只保存在浏览器本地，不会随账号同步。
-- AI API Key 目前保存在浏览器本地，并由浏览器直接请求配置的 OpenAI 兼容接口。
+- 本地直连模式下，AI API Key 保存在浏览器本地，并由浏览器直接请求配置的 OpenAI 兼容接口；只适合本地或个人使用，公网多人使用不安全。公网部署推荐使用后台渠道，把真实密钥保存在服务端配置中。
 - 服务器素材库目前主要保存 URL 或文本，暂未提供文件上传接口。
+- Seedance 本地参考图/视频上传依赖 `PUBLIC_BASE_URL`，如果服务部署在 localhost、内网或不可被火山访问的地址，火山无法拉取参考素材。
+- Seedance 返回远程视频 URL 时，前端会尽量下载为本地 Blob 持久化；如果因 CORS 或网络限制无法下载，会保留远程 URL，后续是否可播放取决于上游 URL 的有效期。
 - 画布更适合桌面端使用，移动端触控体验还未系统完善。
diff --git a/docs/pending-test.md b/docs/pending-test.md
index 5867f38..3b1a416 100644
--- a/docs/pending-test.md
+++ b/docs/pending-test.md
@@ -21,3 +21,6 @@
 - 视频生成前端会识别后端 `{ code, msg }` 错误响应，创建失败不再继续轮询 `undefined`。
 - 新增 `/video` 视频创作台页面，参考生图工作台布局，支持提示词、参考图、视频参数、生成结果、保存素材、下载和本地生成记录；清晰度和秒数均支持常用值选择与手动输入，生成记录只保存媒体 `storageKey` 并可回填本次提示词、参考图和参数。
 - 视频创作台生成前会把模型、尺寸、秒数、清晰度归一化为视频接口支持的参数，并展示后端返回的错误信息，避免页面侧残留的生图参数影响视频请求。
+- 火山方舟 Agent Plan / Seedance 2.0 视频生成需要在真实账号下验证：`/contents/generations/tasks` 创建任务、轮询状态、`content.video_url` 回填画布，以及 401/403/429/超时错误提示。
+- `PUBLIC_BASE_URL` 已配置公网域名时，需要验证本地上传参考视频能被火山拉取；未配置或配置为内网地址时，需要验证前端能给出明确提示。
+- Seedance 返回远程视频 URL 但浏览器无法下载为 Blob 时，需要验证视频节点刷新后仍保留远程 URL，并确认上游 URL 有效期限制。

From 1b557a86d72b9ab8489a026fb241c4cab7aabf67 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 10:01:15 +0800
Subject: [PATCH 05/14] fix: skip text test for ark seedance channels

---
 service/settings.go | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/service/settings.go b/service/settings.go
index 23b2913..ef34edf 100644
--- a/service/settings.go
+++ b/service/settings.go
@@ -44,6 +44,9 @@ func AdminChannelModels(index *int, channel model.ModelChannel) ([]string, error
 	if err != nil {
 		return nil, err
 	}
+	if isArkAgentPlanChannel(resolved) {
+		return []string{"doubao-seedance-2.0"}, nil
+	}
 	return fetchAdminChannelModels(resolved)
 }
 
@@ -52,6 +55,9 @@ func AdminTestChannelModel(index *int, channel model.ModelChannel, modelName str
 	if err != nil {
 		return "", err
 	}
+	if isArkAgentPlanChannel(resolved) || isSeedanceModelName(modelName) {
+		return testArkSeedanceChannelModel(resolved, modelName)
+	}
 	return testAdminChannelModel(resolved, modelName)
 }
 
@@ -186,6 +192,16 @@ func BuildModelChannelURL(channel model.ModelChannel, path string) string {
 	return baseURL + path
 }
 
+func isArkAgentPlanChannel(channel model.ModelChannel) bool {
+	baseURL := strings.TrimRight(strings.ToLower(strings.TrimSpace(channel.BaseURL)), "/")
+	return strings.HasSuffix(baseURL, "/api/plan/v3")
+}
+
+func isSeedanceModelName(modelName string) bool {
+	modelName = strings.ToLower(strings.TrimSpace(modelName))
+	return strings.Contains(modelName, "seedance") || strings.Contains(modelName, "doubao-seedance")
+}
+
 func normalizeModelChannel(channel model.ModelChannel) model.ModelChannel {
 	if channel.Protocol == "" {
 		channel.Protocol = "openai"
@@ -304,6 +320,22 @@ func testAdminChannelModel(channel model.ModelChannel, modelName string) (string
 	return "ok", nil
 }
 
+func testArkSeedanceChannelModel(channel model.ModelChannel, modelName string) (string, error) {
+	if strings.TrimSpace(modelName) == "" {
+		return "", errors.New("缺少模型名称")
+	}
+	if strings.TrimSpace(channel.BaseURL) == "" {
+		return "", safeMessageError{message: "缺少接口地址"}
+	}
+	if strings.TrimSpace(channel.APIKey) == "" {
+		return "", safeMessageError{message: "缺少 API Key"}
+	}
+	if !isArkAgentPlanChannel(channel) {
+		return "", safeMessageError{message: "Seedance 2.0 请使用火山方舟 Agent Plan Base URL：https://ark.cn-beijing.volces.com/api/plan/v3"}
+	}
+	return "Agent Plan / Seedance 视频模型配置格式已通过。后台测试不会调用视频生成接口，请在画布中使用视频生成验证模型权限。", nil
+}
+
 func readAdminChannelError(body []byte, statusCode int, fallback string) error {
 	var payload struct {
 		Error *struct {

From fa9a3c1920aaa26349b9d01619308d54b310e913 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 10:08:01 +0800
Subject: [PATCH 06/14] fix: normalize ark agent plan base url

---
 service/settings.go                | 13 +++++++++++--
 web/src/stores/use-config-store.ts |  4 +++-
 2 files changed, 14 insertions(+), 3 deletions(-)

diff --git a/service/settings.go b/service/settings.go
index ef34edf..e84e5c1 100644
--- a/service/settings.go
+++ b/service/settings.go
@@ -185,15 +185,24 @@ func SelectModelChannel(modelName string) (model.ModelChannel, error) {
 }
 
 func BuildModelChannelURL(channel model.ModelChannel, path string) string {
-	baseURL := strings.TrimRight(channel.BaseURL, "/")
+	baseURL := normalizeModelChannelBaseURL(channel.BaseURL)
 	if !strings.HasSuffix(baseURL, "/v1") && !strings.HasSuffix(baseURL, "/api/v3") && !strings.HasSuffix(baseURL, "/api/plan/v3") {
 		baseURL += "/v1"
 	}
 	return baseURL + path
 }
 
+func normalizeModelChannelBaseURL(baseURL string) string {
+	baseURL = strings.TrimRight(strings.TrimSpace(baseURL), "/")
+	lowerBaseURL := strings.ToLower(baseURL)
+	if index := strings.Index(lowerBaseURL, "/api/plan/v3"); index >= 0 {
+		return baseURL[:index+len("/api/plan/v3")]
+	}
+	return baseURL
+}
+
 func isArkAgentPlanChannel(channel model.ModelChannel) bool {
-	baseURL := strings.TrimRight(strings.ToLower(strings.TrimSpace(channel.BaseURL)), "/")
+	baseURL := strings.ToLower(normalizeModelChannelBaseURL(channel.BaseURL))
 	return strings.HasSuffix(baseURL, "/api/plan/v3")
 }
 
diff --git a/web/src/stores/use-config-store.ts b/web/src/stores/use-config-store.ts
index 08f2912..0bb7296 100644
--- a/web/src/stores/use-config-store.ts
+++ b/web/src/stores/use-config-store.ts
@@ -125,7 +125,9 @@ export function useEffectiveConfig() {
 }
 
 export function buildApiUrl(baseUrl: string, path: string) {
-    const normalizedBaseUrl = baseUrl.trim().replace(/\/+$/, "");
+    let normalizedBaseUrl = baseUrl.trim().replace(/\/+$/, "");
+    const arkPlanIndex = normalizedBaseUrl.toLowerCase().indexOf("/api/plan/v3");
+    if (arkPlanIndex >= 0) normalizedBaseUrl = normalizedBaseUrl.slice(0, arkPlanIndex + "/api/plan/v3".length);
     const apiBaseUrl = normalizedBaseUrl.endsWith("/v1") || normalizedBaseUrl.endsWith("/api/v3") || normalizedBaseUrl.endsWith("/api/plan/v3") ? normalizedBaseUrl : `${normalizedBaseUrl}/v1`;
     return `${apiBaseUrl}${path}`;
 }

From 3be0df0df67adb1c5da5d8b5c33992fe1ecd4d13 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 10:52:59 +0800
Subject: [PATCH 07/14] fix: clarify ark agent plan model discovery

---
 docs/canvas-node-manual.md                  |  3 +-
 docs/features.md                            |  4 +-
 service/settings.go                         | 38 +++++++++----
 service/settings_test.go                    | 62 +++++++++++++++++++++
 web/src/app/(admin)/admin/settings/page.tsx |  7 ++-
 web/src/stores/use-config-store.ts          | 24 +++++++-
 6 files changed, 121 insertions(+), 17 deletions(-)
 create mode 100644 service/settings_test.go

diff --git a/docs/canvas-node-manual.md b/docs/canvas-node-manual.md
index 48db9b8..4849bac 100644
--- a/docs/canvas-node-manual.md
+++ b/docs/canvas-node-manual.md
@@ -20,7 +20,7 @@
 - 当文本节点已有内容时，输入框用于填写想把本段文本修改成什么；点击发送后，会在右侧生成新的文本节点，并自动连接原节点和新节点。
 - 输入内容可以手写，也可以从提示词库选择。
 - 对话框里的模型下拉来自全局配置里已拉取的模型列表；选择结果只作用于当前节点，不会修改其它节点或全局默认模型。
-- 如果下拉中没有模型，需要先打开配置弹窗拉取模型列表，并设置默认生图模型和默认文本模型。
+- 如果下拉中没有模型，需要先打开配置弹窗拉取模型列表，并设置默认生图模型和默认文本模型。火山方舟 Agent Plan 若提示不支持 `/models`，请手动填写模型名。
 
 ### 用文本节点生成图片
 
@@ -48,6 +48,7 @@
 - 生成配置节点的视频模式会读取上游文本作为 prompt，读取上游图片作为参考图，读取上游视频作为参考视频，并在输入预览里显示参考视频。
 - 视频生成接口支持 OpenAI 风格的 `POST /v1/videos`、`GET /v1/videos/{id}` 和 `GET /v1/videos/{id}/content`。
 - 使用火山方舟 Agent Plan / Seedance 2.0 时，Base URL 配置为 `https://ark.cn-beijing.volces.com/api/plan/v3`，模型名使用 Seedance 2.0 对应模型；系统会改用 `POST /contents/generations/tasks` 创建异步任务，并轮询 `GET /contents/generations/tasks/{id}`。
+- Agent Plan 专属 `/api/plan/v3` 当前未提供 OpenAI `/models` 模型列表接口，后台不会伪造模型列表；请手动填写 `doubao-seedance-2.0` 或文档列出的其他可用模型。
 - Seedance 参考视频必须是公网可访问 URL，或由本项目后端在配置 `PUBLIC_BASE_URL` 后上传并暴露的参考素材 URL。本地/内网地址无法被火山服务器拉取。
 
 ### 推荐流程
diff --git a/docs/features.md b/docs/features.md
index 344dd90..5b7a1ca 100644
--- a/docs/features.md
+++ b/docs/features.md
@@ -62,7 +62,7 @@ OpenAI 兼容图像和文本能力继续复用现有接口：
 - `/v1/images/generations`：文生图。
 - `/v1/images/edits`：图生图/参考图编辑。
 - `/v1/chat/completions`：文本问答和带图问答。
-- `/v1/models`：读取模型列表。
+- `/v1/models`：读取模型列表；火山方舟 Agent Plan 专属 `/api/plan/v3` 当前未提供 OpenAI `/models` 模型列表接口，需要手动填写模型名。
 
 视频能力支持两类接口：
 
@@ -71,6 +71,8 @@ OpenAI 兼容图像和文本能力继续复用现有接口：
 
 Base URL 如果已经以 `/v1`、`/api/v3` 或 `/api/plan/v3` 结尾，系统不会再追加 `/v1`。因此 cpa 反代或火山方舟 Agent Plan 可以继续通过现有 Base URL + API Key + Model 方式配置，不需要新增火山生图 Provider。
 
+后台“拉取模型列表”会尝试真实请求 OpenAI `/models`，不会为 Agent Plan 伪造模型结果。如果火山方舟 Agent Plan 返回 404，请手动增加 `doubao-seedance-2.0` 或文档列出的其他模型名。
+
 可配置项：
 
 - Base URL。
diff --git a/service/settings.go b/service/settings.go
index e84e5c1..4a943ab 100644
--- a/service/settings.go
+++ b/service/settings.go
@@ -7,6 +7,7 @@ import (
 	"io"
 	"math/rand"
 	"net/http"
+	"net/url"
 	"sort"
 	"strings"
 
@@ -44,9 +45,6 @@ func AdminChannelModels(index *int, channel model.ModelChannel) ([]string, error
 	if err != nil {
 		return nil, err
 	}
-	if isArkAgentPlanChannel(resolved) {
-		return []string{"doubao-seedance-2.0"}, nil
-	}
 	return fetchAdminChannelModels(resolved)
 }
 
@@ -186,7 +184,8 @@ func SelectModelChannel(modelName string) (model.ModelChannel, error) {
 
 func BuildModelChannelURL(channel model.ModelChannel, path string) string {
 	baseURL := normalizeModelChannelBaseURL(channel.BaseURL)
-	if !strings.HasSuffix(baseURL, "/v1") && !strings.HasSuffix(baseURL, "/api/v3") && !strings.HasSuffix(baseURL, "/api/plan/v3") {
+	lowerBaseURL := strings.ToLower(baseURL)
+	if !strings.HasSuffix(lowerBaseURL, "/v1") && !strings.HasSuffix(lowerBaseURL, "/api/v3") && !strings.HasSuffix(lowerBaseURL, "/api/plan/v3") {
 		baseURL += "/v1"
 	}
 	return baseURL + path
@@ -194,9 +193,20 @@ func BuildModelChannelURL(channel model.ModelChannel, path string) string {
 
 func normalizeModelChannelBaseURL(baseURL string) string {
 	baseURL = strings.TrimRight(strings.TrimSpace(baseURL), "/")
-	lowerBaseURL := strings.ToLower(baseURL)
-	if index := strings.Index(lowerBaseURL, "/api/plan/v3"); index >= 0 {
-		return baseURL[:index+len("/api/plan/v3")]
+	parsed, err := url.Parse(baseURL)
+	if err == nil && parsed.Scheme != "" && parsed.Host != "" {
+		path := strings.TrimRight(parsed.Path, "/")
+		lowerPath := strings.ToLower(path)
+		if index := strings.Index(lowerPath, "/api/plan/v3"); index >= 0 {
+			end := index + len("/api/plan/v3")
+			if len(lowerPath) == end || lowerPath[end] == '/' {
+				parsed.Path = path[:end]
+				parsed.RawPath = ""
+				parsed.RawQuery = ""
+				parsed.Fragment = ""
+				return strings.TrimRight(parsed.String(), "/")
+			}
+		}
 	}
 	return baseURL
 }
@@ -271,6 +281,9 @@ func fetchAdminChannelModels(channel model.ModelChannel) ([]string, error) {
 	defer response.Body.Close()
 	body, _ := io.ReadAll(response.Body)
 	if response.StatusCode >= http.StatusBadRequest {
+		if response.StatusCode == http.StatusNotFound && isArkAgentPlanChannel(channel) {
+			return nil, safeMessageError{message: "火山方舟 Agent Plan 未提供 OpenAI /models 模型列表接口，请手动填写模型名称，例如 doubao-seedance-2.0。"}
+		}
 		return nil, readAdminChannelError(body, response.StatusCode, "读取模型失败")
 	}
 	var payload struct {
@@ -340,9 +353,9 @@ func testArkSeedanceChannelModel(channel model.ModelChannel, modelName string) (
 		return "", safeMessageError{message: "缺少 API Key"}
 	}
 	if !isArkAgentPlanChannel(channel) {
-		return "", safeMessageError{message: "Seedance 2.0 请使用火山方舟 Agent Plan Base URL：https://ark.cn-beijing.volces.com/api/plan/v3"}
+		return "Seedance 视频模型不会发送 /chat/completions 文本测试。已检查 Base URL、API Key 和模型名非空；未调用视频生成接口，因此未验证套餐额度或模型权限。", nil
 	}
-	return "Agent Plan / Seedance 视频模型配置格式已通过。后台测试不会调用视频生成接口，请在画布中使用视频生成验证模型权限。", nil
+	return "Agent Plan / Seedance 视频模型配置格式已通过。后台测试不会调用视频生成接口，因此未验证 API Key、套餐额度或模型权限；请在画布中使用视频生成验证。", nil
 }
 
 func readAdminChannelError(body []byte, statusCode int, fallback string) error {
@@ -360,8 +373,11 @@ func readAdminChannelError(body []byte, statusCode int, fallback string) error {
 			return safeMessageError{message: payload.Msg}
 		}
 	}
-	if statusCode == http.StatusUnauthorized {
-		return safeMessageError{message: "上游接口认证失败（401），请检查 API Key"}
+	if statusCode == http.StatusUnauthorized || statusCode == http.StatusForbidden {
+		return safeMessageError{message: fmt.Sprintf("上游接口鉴权失败（%d），请检查 API Key、套餐权限或模型权限", statusCode)}
+	}
+	if statusCode == http.StatusTooManyRequests {
+		return safeMessageError{message: "上游接口限流或额度不足（429），请稍后重试或检查额度"}
 	}
 	if statusCode > 0 {
 		return safeMessageError{message: fmt.Sprintf("%s：%d", fallback, statusCode)}
diff --git a/service/settings_test.go b/service/settings_test.go
new file mode 100644
index 0000000..ea92d83
--- /dev/null
+++ b/service/settings_test.go
@@ -0,0 +1,62 @@
+package service
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/basketikun/infinite-canvas/model"
+)
+
+func TestFetchAdminChannelModelsParsesOpenAIModels(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/models" {
+			t.Fatalf("unexpected path: %s", r.URL.Path)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"data":[{"id":"z-model"},{"id":"a-model"},{"id":""}]}`))
+	}))
+	defer server.Close()
+
+	models, err := fetchAdminChannelModels(model.ModelChannel{
+		BaseURL: server.URL,
+		APIKey:  "test-key",
+	})
+	if err != nil {
+		t.Fatalf("fetchAdminChannelModels returned error: %v", err)
+	}
+	if want := []string{"a-model", "z-model"}; !reflect.DeepEqual(models, want) {
+		t.Fatalf("models = %#v, want %#v", models, want)
+	}
+}
+
+func TestFetchAdminChannelModelsReportsArkPlanModelsUnsupported(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/api/plan/v3/models" {
+			t.Fatalf("unexpected path: %s", r.URL.Path)
+		}
+		http.NotFound(w, r)
+	}))
+	defer server.Close()
+
+	_, err := fetchAdminChannelModels(model.ModelChannel{
+		BaseURL: server.URL + "/api/plan/v3/contents/generations/tasks",
+		APIKey:  "test-key",
+	})
+	if err == nil {
+		t.Fatal("expected unsupported /models error")
+	}
+	if !strings.Contains(err.Error(), "Agent Plan 未提供 OpenAI /models") {
+		t.Fatalf("error = %q", err.Error())
+	}
+}
+
+func TestBuildModelChannelURLNormalizesArkPlanTaskPath(t *testing.T) {
+	got := BuildModelChannelURL(model.ModelChannel{BaseURL: "https://ark.cn-beijing.volces.com/api/plan/v3/contents/generations/tasks?debug=1"}, "/models")
+	want := "https://ark.cn-beijing.volces.com/api/plan/v3/models"
+	if got != want {
+		t.Fatalf("BuildModelChannelURL = %q, want %q", got, want)
+	}
+}
diff --git a/web/src/app/(admin)/admin/settings/page.tsx b/web/src/app/(admin)/admin/settings/page.tsx
index 7ca449e..c262680 100644
--- a/web/src/app/(admin)/admin/settings/page.tsx
+++ b/web/src/app/(admin)/admin/settings/page.tsx
@@ -215,6 +215,10 @@ export default function AdminSettingsPage() {
             const channelModels = await fetchChannelModels(token, { index: editingChannelIndex ?? undefined, channel: normalizeChannel(channel) });
             const current = isModelSelectorOpen ? uniqueModels(modelSelectSelected) : uniqueModels(channelForm.getFieldValue("models") || []);
             rememberModels(channelModels);
+            if (!channelModels.length) {
+                message.warning("上游未返回模型列表，请手动输入模型名称");
+                return;
+            }
             setModelSelectExisting(current);
             setModelSelectSource(uniqueModels(channelModels));
             setModelSelectSelected(uniqueModels([...current, ...channelModels]));
@@ -707,6 +711,7 @@ export default function AdminSettingsPage() {
                                 </Button>
                             </Space.Compact>
                         </Flex>
+                        <Typography.Text type="secondary">如果上游不提供 OpenAI /models 模型列表接口，请在这里手动增加模型名称。</Typography.Text>
                         <Tabs
                             activeKey={modelSelectTab}
                             onChange={(key) => setModelSelectTab(key as ModelSelectTabKey)}
@@ -765,7 +770,7 @@ export default function AdminSettingsPage() {
                     destroyOnHidden
                 >
                     <Flex vertical gap={12}>
-                        <Typography.Text type="secondary">测试会向选中模型发送一条 hi，用于确认渠道是否有响应。</Typography.Text>
+                        <Typography.Text type="secondary">普通文本模型会发送一条 hi；Agent Plan / Seedance 视频模型只做配置格式检查，不会发起视频生成，也不代表模型权限已验证。</Typography.Text>
                         <Input.Search placeholder="搜索模型..." allowClear value={testKeyword} onChange={(event) => setTestKeyword(event.target.value)} />
                         <Table
                             rowKey="model"
diff --git a/web/src/stores/use-config-store.ts b/web/src/stores/use-config-store.ts
index 0bb7296..49e0af4 100644
--- a/web/src/stores/use-config-store.ts
+++ b/web/src/stores/use-config-store.ts
@@ -126,8 +126,26 @@ export function useEffectiveConfig() {
 
 export function buildApiUrl(baseUrl: string, path: string) {
     let normalizedBaseUrl = baseUrl.trim().replace(/\/+$/, "");
-    const arkPlanIndex = normalizedBaseUrl.toLowerCase().indexOf("/api/plan/v3");
-    if (arkPlanIndex >= 0) normalizedBaseUrl = normalizedBaseUrl.slice(0, arkPlanIndex + "/api/plan/v3".length);
-    const apiBaseUrl = normalizedBaseUrl.endsWith("/v1") || normalizedBaseUrl.endsWith("/api/v3") || normalizedBaseUrl.endsWith("/api/plan/v3") ? normalizedBaseUrl : `${normalizedBaseUrl}/v1`;
+    normalizedBaseUrl = normalizeArkPlanBaseUrl(normalizedBaseUrl);
+    const lowerBaseUrl = normalizedBaseUrl.toLowerCase();
+    const apiBaseUrl = lowerBaseUrl.endsWith("/v1") || lowerBaseUrl.endsWith("/api/v3") || lowerBaseUrl.endsWith("/api/plan/v3") ? normalizedBaseUrl : `${normalizedBaseUrl}/v1`;
     return `${apiBaseUrl}${path}`;
 }
+
+function normalizeArkPlanBaseUrl(baseUrl: string) {
+    try {
+        const url = new URL(baseUrl);
+        const path = url.pathname.replace(/\/+$/, "");
+        const lowerPath = path.toLowerCase();
+        const arkPlanIndex = lowerPath.indexOf("/api/plan/v3");
+        if (arkPlanIndex < 0) return baseUrl;
+        const end = arkPlanIndex + "/api/plan/v3".length;
+        if (lowerPath.length !== end && lowerPath[end] !== "/") return baseUrl;
+        url.pathname = path.slice(0, end);
+        url.search = "";
+        url.hash = "";
+        return url.toString().replace(/\/+$/, "");
+    } catch {
+        return baseUrl;
+    }
+}

From 871bedc07239b57e96ee9531d5661645575cbc50 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 11:00:20 +0800
Subject: [PATCH 08/14] fix: add timeout for admin channel probes

---
 service/settings.go | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/service/settings.go b/service/settings.go
index 4a943ab..51ac403 100644
--- a/service/settings.go
+++ b/service/settings.go
@@ -10,11 +10,14 @@ import (
 	"net/url"
 	"sort"
 	"strings"
+	"time"
 
 	"github.com/basketikun/infinite-canvas/model"
 	"github.com/basketikun/infinite-canvas/repository"
 )
 
+var adminModelHTTPClient = &http.Client{Timeout: 30 * time.Second}
+
 func PublicSettings() (model.PublicSetting, error) {
 	settings, err := repository.GetSettings()
 	return normalizePublicSetting(settings.Public), err
@@ -274,9 +277,9 @@ func fetchAdminChannelModels(channel model.ModelChannel) ([]string, error) {
 		return nil, err
 	}
 	request.Header.Set("Authorization", "Bearer "+channel.APIKey)
-	response, err := http.DefaultClient.Do(request)
+	response, err := adminModelHTTPClient.Do(request)
 	if err != nil {
-		return nil, err
+		return nil, safeMessageError{message: "读取模型失败：上游接口无响应或网络不可达"}
 	}
 	defer response.Body.Close()
 	body, _ := io.ReadAll(response.Body)
@@ -319,9 +322,9 @@ func testAdminChannelModel(channel model.ModelChannel, modelName string) (string
 	}
 	request.Header.Set("Authorization", "Bearer "+channel.APIKey)
 	request.Header.Set("Content-Type", "application/json")
-	response, err := http.DefaultClient.Do(request)
+	response, err := adminModelHTTPClient.Do(request)
 	if err != nil {
-		return "", err
+		return "", safeMessageError{message: "测试失败：上游接口无响应或网络不可达"}
 	}
 	defer response.Body.Close()
 	responseBody, _ := io.ReadAll(response.Body)

From 4bffb55695df2ff43d0487b51f7a44b4e7cc741f Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 13:02:36 +0800
Subject: [PATCH 09/14] feat: add seedance media params and model visibility

---
 docs/features.md                              |   4 +-
 docs/pending-test.md                          |   6 +-
 docs/system-settings.md                       |   9 +-
 handler/media_reference.go                    |  58 +++++++-
 handler/media_reference_test.go               |  40 ++++++
 service/settings.go                           |  75 +++++++++-
 service/settings_test.go                      |  38 +++++
 web/src/app/(admin)/admin/settings/page.tsx   |  11 +-
 .../(user)/canvas/[id]/canvas-client-page.tsx |  97 +++++++++----
 .../components/canvas-config-node-panel.tsx   |  58 +++++++-
 .../canvas/components/canvas-mini-map.tsx     |   2 +-
 .../components/canvas-node-generation.ts      |  24 +++-
 .../components/canvas-node-hover-toolbar.tsx  |  11 +-
 .../components/canvas-node-prompt-panel.tsx   |  11 +-
 .../(user)/canvas/components/canvas-node.tsx  |  25 +++-
 .../canvas/components/canvas-toolbar.tsx      |  12 +-
 web/src/app/(user)/canvas/constants.ts        |   5 +
 web/src/app/(user)/canvas/types.ts            |   4 +
 web/src/app/(user)/video/page.tsx             | 130 +++++++++++++++---
 web/src/components/video-settings-panel.tsx   | 103 +++++++++++++-
 web/src/lib/seedance-video.ts                 | 125 +++++++++++++++++
 web/src/services/api/video.ts                 |  98 +++++++------
 web/src/services/file-storage.ts              |  18 ++-
 web/src/stores/use-config-store.ts            |  39 +++++-
 web/src/types/media.ts                        |  10 ++
 25 files changed, 873 insertions(+), 140 deletions(-)
 create mode 100644 handler/media_reference_test.go
 create mode 100644 web/src/lib/seedance-video.ts

diff --git a/docs/features.md b/docs/features.md
index 5b7a1ca..a6963d0 100644
--- a/docs/features.md
+++ b/docs/features.md
@@ -82,11 +82,11 @@ Base URL 如果已经以 `/v1`、`/api/v3` 或 `/api/plan/v3` 结尾，系统不
 - 图片比例。
 - 生成数量。
 - 视频模型。
-- 视频比例、清晰度和时长。
+- 视频比例、清晰度、时长、生成声音和水印。
 
 普通图片/文本节点可以直接输入提示词生成结果。生成配置节点可以读取上游节点内容，并按节点自己的配置批量生成多个图片或文本结果。生成配置节点支持预览当前提示词和参考图输入，并调整输入顺序。
 
-视频生成可从文本节点读取 prompt，从图片节点读取参考图，从视频节点读取参考视频。生成成功后会把视频插入画布为视频节点并使用原生播放器预览。Seedance 参考视频需要公网可访问 URL；本地上传视频会先通过 `/api/v1/media/references` 保存到服务端，再由 `PUBLIC_BASE_URL` 生成可供火山服务器拉取的公开链接。
+视频生成可从文本节点读取 prompt，从图片节点读取参考图，从视频节点读取参考视频，从音频节点读取参考音频。Seedance 2.0 支持最多 9 张参考图、3 个参考视频、3 个参考音频；分辨率支持 `480p`、`720p`、`1080p`（fast 模型不支持 `1080p`），比例支持 `16:9`、`4:3`、`1:1`、`3:4`、`9:16`、`21:9`、`adaptive`，时长支持 4-15 秒或智能时长。生成成功后会把视频插入画布为视频节点并使用原生播放器预览。Seedance 参考视频和参考音频需要公网可访问 URL；本地上传素材会先通过 `/api/v1/media/references` 保存到服务端，再由 `PUBLIC_BASE_URL` 生成可供火山服务器拉取的公开链接。
 
 ## 画布助手
 
diff --git a/docs/pending-test.md b/docs/pending-test.md
index 3b1a416..e313d2b 100644
--- a/docs/pending-test.md
+++ b/docs/pending-test.md
@@ -22,5 +22,9 @@
 - 新增 `/video` 视频创作台页面，参考生图工作台布局，支持提示词、参考图、视频参数、生成结果、保存素材、下载和本地生成记录；清晰度和秒数均支持常用值选择与手动输入，生成记录只保存媒体 `storageKey` 并可回填本次提示词、参考图和参数。
 - 视频创作台生成前会把模型、尺寸、秒数、清晰度归一化为视频接口支持的参数，并展示后端返回的错误信息，避免页面侧残留的生图参数影响视频请求。
 - 火山方舟 Agent Plan / Seedance 2.0 视频生成需要在真实账号下验证：`/contents/generations/tasks` 创建任务、轮询状态、`content.video_url` 回填画布，以及 401/403/429/超时错误提示。
-- `PUBLIC_BASE_URL` 已配置公网域名时，需要验证本地上传参考视频能被火山拉取；未配置或配置为内网地址时，需要验证前端能给出明确提示。
+- 管理后台保存私有渠道后，需要验证所有已启用渠道里的模型会自动出现在公开 `availableModels`，并且 `defaultVideoModel`、`defaultImageModel`、`defaultTextModel` 在为空或失效时会自动修复，前台不再显示旧的 `grok` 默认值。
+- `/video` 和画布视频设置已按 Seedance 2.0 增加分辨率、比例、4-15 秒/智能时长、生成声音和水印参数；需要在真实浏览器里验证参数回填、生成记录和画布节点配置都能保持一致。
+- `/video` 支持最多 9 张参考图、3 个参考视频、3 个参考音频；需要验证格式、大小、音频时长提示和生成请求中的 `reference_image`、`reference_video`、`reference_audio` 组装。
+- 画布新增音频节点，支持上传、拖入、播放、移动、缩放、删除，并可作为上游参考音频参与 Seedance 视频生成；需要验证刷新后本地音频 URL 能恢复。
+- `PUBLIC_BASE_URL` 已配置公网域名时，需要验证本地上传参考视频和参考音频能被火山拉取；未配置或配置为内网地址时，需要验证前端能给出明确提示。
 - Seedance 返回远程视频 URL 但浏览器无法下载为 Blob 时，需要验证视频节点刷新后仍保留远程 URL，并确认上游 URL 有效期限制。
diff --git a/docs/system-settings.md b/docs/system-settings.md
index 74a267b..0996665 100644
--- a/docs/system-settings.md
+++ b/docs/system-settings.md
@@ -41,11 +41,12 @@
 
 | 字段 | 类型 | 说明 |
 | --- | --- | --- |
-| `availableModels` | string[] | 系统可用模型，由管理员手动选择；页面下拉选项可来自私有渠道模型 |
+| `availableModels` | string[] | 系统可用模型；保存设置时会自动合并所有已启用私有渠道的模型 |
 | `modelCosts` | object[] | 模型算力点配置，后端模型接口调用前按模型预扣，上游失败时返还；未配置默认不扣除 |
-| `defaultModel` | string | 默认模型，从 `availableModels` 中选择 |
-| `defaultImageModel` | string | 默认图片模型，从 `availableModels` 中选择 |
-| `defaultTextModel` | string | 默认文本模型，从 `availableModels` 中选择 |
+| `defaultModel` | string | 默认模型，从 `availableModels` 中选择；为空或失效时优先选择文本模型 |
+| `defaultImageModel` | string | 默认图片模型，从 `availableModels` 中选择；为空或失效时优先选择 `seedream`、`image`、`gpt-image` 模型 |
+| `defaultVideoModel` | string | 默认视频模型，从 `availableModels` 中选择；为空或失效时优先选择 `seedance`、`video` 模型 |
+| `defaultTextModel` | string | 默认文本模型，从 `availableModels` 中选择；为空或失效时优先选择非图片/视频模型 |
 | `systemPrompt` | string | 系统提示词 |
 | `allowCustomChannel` | boolean | 是否允许用户在配置弹窗中切换为本地直连渠道，默认允许 |
 
diff --git a/handler/media_reference.go b/handler/media_reference.go
index d080f97..b528588 100644
--- a/handler/media_reference.go
+++ b/handler/media_reference.go
@@ -12,7 +12,16 @@ import (
 	"github.com/google/uuid"
 )
 
-const referenceMediaMaxBytes = 80 << 20
+const (
+	referenceMediaMaxBytes    = 80 << 20
+	referenceImageMaxBytes    = 30 << 20
+	referenceVideoMaxBytes    = 50 << 20
+	referenceAudioMaxBytes    = 15 << 20
+	referenceImageAllowedText = "jpeg/png/webp/bmp/gif/heic/heif 图片"
+	referenceVideoAllowedText = "mp4/mov 视频"
+	referenceAudioAllowedText = "mp3/wav 音频"
+	referenceMediaAllowedText = referenceImageAllowedText + "、" + referenceVideoAllowedText + "或" + referenceAudioAllowedText
+)
 
 type referenceMediaUploadResult struct {
 	ID       string `json:"id"`
@@ -24,7 +33,7 @@ type referenceMediaUploadResult struct {
 func UploadReferenceMedia(w http.ResponseWriter, r *http.Request) {
 	publicBaseURL := strings.TrimRight(strings.TrimSpace(config.Cfg.PublicBaseURL), "/")
 	if publicBaseURL == "" {
-		Fail(w, "未配置 PUBLIC_BASE_URL，无法把本地参考视频提供给火山方舟访问")
+		Fail(w, "未配置 PUBLIC_BASE_URL，无法把本地参考素材提供给火山方舟访问")
 		return
 	}
 	r.Body = http.MaxBytesReader(w, r.Body, referenceMediaMaxBytes+1)
@@ -44,7 +53,7 @@ func UploadReferenceMedia(w http.ResponseWriter, r *http.Request) {
 
 	mimeType, ext, ok := normalizeReferenceMediaType(header.Header.Get("Content-Type"), filepath.Ext(header.Filename))
 	if !ok {
-		Fail(w, "参考素材格式不支持，请使用 jpeg/png/webp/bmp/gif/heic/heif 图片或 mp4/mov 视频")
+		Fail(w, "参考素材格式不支持，请使用 "+referenceMediaAllowedText)
 		return
 	}
 	if err := os.MkdirAll(referenceMediaDir(), 0o755); err != nil {
@@ -65,9 +74,14 @@ func UploadReferenceMedia(w http.ResponseWriter, r *http.Request) {
 		Fail(w, "参考素材保存失败")
 		return
 	}
-	if bytes <= 0 || bytes > referenceMediaMaxBytes {
+	if bytes <= 0 {
 		_ = os.Remove(targetPath)
-		Fail(w, "参考素材为空或超过大小限制")
+		Fail(w, "参考素材为空")
+		return
+	}
+	if limit := referenceMediaTypeMaxBytes(mimeType); limit > 0 && bytes > limit {
+		_ = os.Remove(targetPath)
+		Fail(w, referenceMediaSizeMessage(mimeType))
 		return
 	}
 	OK(w, referenceMediaUploadResult{
@@ -141,6 +155,10 @@ func referenceMediaExtByMimeType(mimeType string) string {
 		return ".mp4"
 	case "video/quicktime", "video/mov":
 		return ".mov"
+	case "audio/mpeg", "audio/mp3":
+		return ".mp3"
+	case "audio/wav", "audio/x-wav", "audio/wave":
+		return ".wav"
 	default:
 		return ""
 	}
@@ -166,7 +184,37 @@ func mimeTypeByReferenceMediaExt(ext string) string {
 		return "video/mp4"
 	case ".mov":
 		return "video/quicktime"
+	case ".mp3":
+		return "audio/mpeg"
+	case ".wav":
+		return "audio/wav"
 	default:
 		return ""
 	}
 }
+
+func referenceMediaTypeMaxBytes(mimeType string) int64 {
+	if strings.HasPrefix(mimeType, "image/") {
+		return referenceImageMaxBytes
+	}
+	if strings.HasPrefix(mimeType, "video/") {
+		return referenceVideoMaxBytes
+	}
+	if strings.HasPrefix(mimeType, "audio/") {
+		return referenceAudioMaxBytes
+	}
+	return referenceMediaMaxBytes
+}
+
+func referenceMediaSizeMessage(mimeType string) string {
+	if strings.HasPrefix(mimeType, "image/") {
+		return "参考图片超过大小限制，请使用 30MB 以内的图片"
+	}
+	if strings.HasPrefix(mimeType, "video/") {
+		return "参考视频超过大小限制，请使用 50MB 以内的 mp4/mov 视频"
+	}
+	if strings.HasPrefix(mimeType, "audio/") {
+		return "参考音频超过大小限制，请使用 15MB 以内的 mp3/wav 音频"
+	}
+	return "参考素材超过大小限制"
+}
diff --git a/handler/media_reference_test.go b/handler/media_reference_test.go
new file mode 100644
index 0000000..95cd7fe
--- /dev/null
+++ b/handler/media_reference_test.go
@@ -0,0 +1,40 @@
+package handler
+
+import "testing"
+
+func TestNormalizeReferenceMediaTypeSupportsAudio(t *testing.T) {
+	tests := []struct {
+		name        string
+		contentType string
+		ext         string
+		wantMime    string
+		wantExt     string
+	}{
+		{name: "mp3 mime", contentType: "audio/mpeg", ext: ".bin", wantMime: "audio/mpeg", wantExt: ".mp3"},
+		{name: "wav ext fallback", contentType: "application/octet-stream", ext: ".wav", wantMime: "audio/wav", wantExt: ".wav"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mimeType, ext, ok := normalizeReferenceMediaType(tt.contentType, tt.ext)
+			if !ok {
+				t.Fatal("expected media type to be accepted")
+			}
+			if mimeType != tt.wantMime || ext != tt.wantExt {
+				t.Fatalf("got (%q, %q), want (%q, %q)", mimeType, ext, tt.wantMime, tt.wantExt)
+			}
+		})
+	}
+}
+
+func TestReferenceMediaTypeMaxBytes(t *testing.T) {
+	if got := referenceMediaTypeMaxBytes("audio/mpeg"); got != referenceAudioMaxBytes {
+		t.Fatalf("audio max bytes = %d, want %d", got, referenceAudioMaxBytes)
+	}
+	if got := referenceMediaTypeMaxBytes("video/mp4"); got != referenceVideoMaxBytes {
+		t.Fatalf("video max bytes = %d, want %d", got, referenceVideoMaxBytes)
+	}
+	if got := referenceMediaTypeMaxBytes("image/png"); got != referenceImageMaxBytes {
+		t.Fatalf("image max bytes = %d, want %d", got, referenceImageMaxBytes)
+	}
+}
diff --git a/service/settings.go b/service/settings.go
index 51ac403..121c802 100644
--- a/service/settings.go
+++ b/service/settings.go
@@ -20,7 +20,7 @@ var adminModelHTTPClient = &http.Client{Timeout: 30 * time.Second}
 
 func PublicSettings() (model.PublicSetting, error) {
 	settings, err := repository.GetSettings()
-	return normalizePublicSetting(settings.Public), err
+	return normalizeSettings(settings).Public, err
 }
 
 func AdminSettings() (model.Settings, error) {
@@ -63,12 +63,16 @@ func AdminTestChannelModel(index *int, channel model.ModelChannel, modelName str
 }
 
 func normalizeSettings(settings model.Settings) model.Settings {
-	settings.Public = normalizePublicSetting(settings.Public)
 	settings.Private = normalizePrivateSetting(settings.Private)
+	settings.Public = normalizePublicSettingWithChannels(settings.Public, settings.Private.Channels)
 	return settings
 }
 
 func normalizePublicSetting(setting model.PublicSetting) model.PublicSetting {
+	return normalizePublicSettingWithChannels(setting, nil)
+}
+
+func normalizePublicSettingWithChannels(setting model.PublicSetting, channels []model.ModelChannel) model.PublicSetting {
 	if setting.ModelChannel.AvailableModels == nil {
 		setting.ModelChannel.AvailableModels = []string{}
 	}
@@ -89,6 +93,16 @@ func normalizePublicSetting(setting model.PublicSetting) model.PublicSetting {
 		enabled := true
 		setting.Auth.AllowRegister = &enabled
 	}
+	enabledModels := enabledChannelModels(channels)
+	if len(enabledModels) > 0 {
+		setting.ModelChannel.AvailableModels = enabledModels
+	} else {
+		setting.ModelChannel.AvailableModels = uniqueModelNames(setting.ModelChannel.AvailableModels)
+	}
+	setting.ModelChannel.DefaultTextModel = repairDefaultModel(setting.ModelChannel.DefaultTextModel, setting.ModelChannel.AvailableModels, isTextModelName)
+	setting.ModelChannel.DefaultImageModel = repairDefaultModel(setting.ModelChannel.DefaultImageModel, setting.ModelChannel.AvailableModels, isImageModelName)
+	setting.ModelChannel.DefaultVideoModel = repairDefaultModel(setting.ModelChannel.DefaultVideoModel, setting.ModelChannel.AvailableModels, isVideoModelName)
+	setting.ModelChannel.DefaultModel = repairDefaultModel(setting.ModelChannel.DefaultModel, setting.ModelChannel.AvailableModels, isTextModelName)
 	return setting
 }
 
@@ -224,6 +238,63 @@ func isSeedanceModelName(modelName string) bool {
 	return strings.Contains(modelName, "seedance") || strings.Contains(modelName, "doubao-seedance")
 }
 
+func enabledChannelModels(channels []model.ModelChannel) []string {
+	models := []string{}
+	for _, channel := range channels {
+		if !channel.Enabled {
+			continue
+		}
+		models = append(models, channel.Models...)
+	}
+	return uniqueModelNames(models)
+}
+
+func uniqueModelNames(models []string) []string {
+	result := []string{}
+	seen := map[string]bool{}
+	for _, item := range models {
+		name := strings.TrimSpace(item)
+		if name == "" || seen[name] {
+			continue
+		}
+		seen[name] = true
+		result = append(result, name)
+	}
+	return result
+}
+
+func repairDefaultModel(current string, models []string, preferred func(string) bool) string {
+	current = strings.TrimSpace(current)
+	for _, item := range models {
+		if item == current {
+			return current
+		}
+	}
+	for _, item := range models {
+		if preferred(item) {
+			return item
+		}
+	}
+	if len(models) > 0 {
+		return models[0]
+	}
+	return ""
+}
+
+func isVideoModelName(modelName string) bool {
+	name := strings.ToLower(strings.TrimSpace(modelName))
+	return strings.Contains(name, "seedance") || strings.Contains(name, "video")
+}
+
+func isImageModelName(modelName string) bool {
+	name := strings.ToLower(strings.TrimSpace(modelName))
+	return strings.Contains(name, "seedream") || strings.Contains(name, "gpt-image") || strings.Contains(name, "image")
+}
+
+func isTextModelName(modelName string) bool {
+	return !isImageModelName(modelName) && !isVideoModelName(modelName)
+}
+
 func normalizeModelChannel(channel model.ModelChannel) model.ModelChannel {
 	if channel.Protocol == "" {
 		channel.Protocol = "openai"
diff --git a/service/settings_test.go b/service/settings_test.go
index ea92d83..27c6b2e 100644
--- a/service/settings_test.go
+++ b/service/settings_test.go
@@ -60,3 +60,41 @@ func TestBuildModelChannelURLNormalizesArkPlanTaskPath(t *testing.T) {
 		t.Fatalf("BuildModelChannelURL = %q, want %q", got, want)
 	}
 }
+
+func TestNormalizeSettingsPublishesEnabledChannelModelsAndRepairsDefaults(t *testing.T) {
+	settings := normalizeSettings(model.Settings{
+		Public: model.PublicSetting{
+			ModelChannel: model.PublicModelChannelSetting{
+				AvailableModels:   []string{"grok-imagine-video", "disabled-model"},
+				DefaultModel:      "grok-imagine-video",
+				DefaultTextModel:  "missing-text",
+				DefaultImageModel: "missing-image",
+				DefaultVideoModel: "missing-video",
+			},
+		},
+		Private: model.PrivateSetting{
+			Channels: []model.ModelChannel{
+				{Enabled: true, Models: []string{"gpt-5.5", "doubao-seedream-5.0-lite", "doubao-seedance-2.0-fast", "gpt-5.5"}},
+				{Enabled: false, Models: []string{"disabled-model"}},
+			},
+		},
+	})
+
+	channel := settings.Public.ModelChannel
+	wantModels := []string{"gpt-5.5", "doubao-seedream-5.0-lite", "doubao-seedance-2.0-fast"}
+	if !reflect.DeepEqual(channel.AvailableModels, wantModels) {
+		t.Fatalf("available models = %#v, want %#v", channel.AvailableModels, wantModels)
+	}
+	if channel.DefaultModel != "gpt-5.5" {
+		t.Fatalf("default model = %q, want text model", channel.DefaultModel)
+	}
+	if channel.DefaultTextModel != "gpt-5.5" {
+		t.Fatalf("default text model = %q, want text model", channel.DefaultTextModel)
+	}
+	if channel.DefaultImageModel != "doubao-seedream-5.0-lite" {
+		t.Fatalf("default image model = %q, want seedream", channel.DefaultImageModel)
+	}
+	if channel.DefaultVideoModel != "doubao-seedance-2.0-fast" {
+		t.Fatalf("default video model = %q, want seedance", channel.DefaultVideoModel)
+	}
+}
diff --git a/web/src/app/(admin)/admin/settings/page.tsx b/web/src/app/(admin)/admin/settings/page.tsx
index c262680..7b5029f 100644
--- a/web/src/app/(admin)/admin/settings/page.tsx
+++ b/web/src/app/(admin)/admin/settings/page.tsx
@@ -341,7 +341,7 @@ export default function AdminSettingsPage() {
         const nextChannelModels = collectChannelModels(nextChannels);
         const nextSettings = normalizeSettings({
             ...values,
-            public: { ...values.public, modelChannel: { ...values.public.modelChannel, availableModels: filterModels(values.public.modelChannel.availableModels, nextChannelModels) } },
+            public: { ...values.public, modelChannel: { ...values.public.modelChannel, availableModels: nextChannelModels } },
             private: { ...values.private, channels: nextChannels },
         });
         const saved = normalizeSettings(await saveAdminSettings(token, nextSettings));
@@ -414,7 +414,7 @@ export default function AdminSettingsPage() {
                             <Form form={form} layout="vertical" initialValues={emptySettings} requiredMark={false}>
                                 <Row gutter={16}>
                                     <Col span={24}>
-                                        <Form.Item name={["public", "modelChannel", "availableModels"]} label="系统可用模型(请先在私有配置里配置渠道)" extra="可选项来自已启用渠道中选择的模型，最终开放哪些模型由这里勾选决定">
+                                        <Form.Item name={["public", "modelChannel", "availableModels"]} label="系统可用模型(请先在私有配置里配置渠道)" extra="保存设置时会自动合并所有已启用私有渠道的模型，前台模型下拉会读取这里的公开列表">
                                             <Select mode="multiple" placeholder="请选择系统可用模型" options={channelModels.map((item) => ({ label: item, value: item }))} />
                                         </Form.Item>
                                     </Col>
@@ -931,11 +931,6 @@ function uniqueModels(models: string[]) {
     return Array.from(new Set(models.filter(Boolean)));
 }
 
-function filterModels(models: string[], options: string[]) {
-    const optionSet = new Set(options);
-    return uniqueModels(models).filter((model) => optionSet.has(model));
-}
-
 function modelSummary(models: string[]) {
     if (!models.length) return "未配置模型";
     const preview = models.slice(0, 3).join(", ");
@@ -971,7 +966,7 @@ async function collectSettings(form: any, editorMode: Record<SettingsTabKey, Edi
         }
         values.private = privateSetting;
     }
-    values.public.modelChannel.availableModels = filterModels(values.public.modelChannel.availableModels, collectChannelModels(values.private.channels));
+    values.public.modelChannel.availableModels = collectChannelModels(values.private.channels);
     return normalizeSettings(values);
 }
 
diff --git a/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx b/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx
index 20020e0..469008b 100644
--- a/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx
+++ b/web/src/app/(user)/canvas/[id]/canvas-client-page.tsx
@@ -3,7 +3,7 @@
 import { useCallback, useEffect, useLayoutEffect, useMemo, useRef, useState } from "react";
 import type { ChangeEvent as ReactChangeEvent, DragEvent as ReactDragEvent, MouseEvent as ReactMouseEvent, PointerEvent as ReactPointerEvent } from "react";
 import { useParams, useRouter } from "next/navigation";
-import { Home, ImageIcon, Images, List, Menu, MessageSquare, Plus, Redo2, Settings2, Trash2, Undo2, Upload, Video } from "lucide-react";
+import { Home, ImageIcon, Images, List, Menu, MessageSquare, Music2, Plus, Redo2, Settings2, Trash2, Undo2, Upload, Video } from "lucide-react";
 import { saveAs } from "file-saver";
 
 import { requestEdit, requestGeneration, requestImageQuestion } from "@/services/api/image";
@@ -52,6 +52,7 @@ import {
     type ViewportTransform,
 } from "../types";
 import type { ReferenceImage } from "@/types/image";
+import type { ReferenceAudio } from "@/types/media";
 
 type CanvasClipboard = {
     nodes: CanvasNodeData[];
@@ -141,7 +142,7 @@ function CanvasRefreshShell() {
     );
 }
 
-function ConnectionCreateMenu({ pending, onCreate, onClose }: { pending: PendingConnectionCreate; onCreate: (type: CanvasNodeType.Image | CanvasNodeType.Text | CanvasNodeType.Config | CanvasNodeType.Video) => void; onClose: () => void }) {
+function ConnectionCreateMenu({ pending, onCreate, onClose }: { pending: PendingConnectionCreate; onCreate: (type: CanvasNodeType.Image | CanvasNodeType.Text | CanvasNodeType.Config | CanvasNodeType.Video | CanvasNodeType.Audio) => void; onClose: () => void }) {
     const theme = canvasThemes[useThemeStore((state) => state.theme)];
     return (
         <div
@@ -163,6 +164,7 @@ function ConnectionCreateMenu({ pending, onCreate, onClose }: { pending: Pending
                 <ConnectionCreateOption theme={theme} icon={<List className="size-5" />} title="文本生成" description="脚本、广告词、品牌文案" onClick={() => onCreate(CanvasNodeType.Text)} />
                 <ConnectionCreateOption theme={theme} icon={<ImageIcon className="size-5" />} title="图片生成" onClick={() => onCreate(CanvasNodeType.Image)} />
                 <ConnectionCreateOption theme={theme} icon={<Video className="size-5" />} title="视频生成" onClick={() => onCreate(CanvasNodeType.Video)} />
+                <ConnectionCreateOption theme={theme} icon={<Music2 className="size-5" />} title="音频参考" onClick={() => onCreate(CanvasNodeType.Audio)} />
                 <ConnectionCreateOption theme={theme} icon={<Settings2 className="size-5" />} title="配置节点" description="模型、尺寸、数量和输入顺序" onClick={() => onCreate(CanvasNodeType.Config)} />
             </div>
         </div>
@@ -483,7 +485,7 @@ function InfiniteCanvasPage() {
     );
 
     const createConnectedNode = useCallback(
-        (type: CanvasNodeType.Image | CanvasNodeType.Text | CanvasNodeType.Config | CanvasNodeType.Video, pending: PendingConnectionCreate) => {
+        (type: CanvasNodeType.Image | CanvasNodeType.Text | CanvasNodeType.Config | CanvasNodeType.Video | CanvasNodeType.Audio, pending: PendingConnectionCreate) => {
             const metadata = type === CanvasNodeType.Config ? { model: effectiveConfig.imageModel || effectiveConfig.model, size: effectiveConfig.size, count: 3 } : undefined;
             const newNode = createCanvasNode(type, pending.position, metadata);
             const connection = normalizeConnection(pending.connection.nodeId, newNode.id, [...nodesRef.current, newNode], pending.connection.handleType);
@@ -495,7 +497,7 @@ function InfiniteCanvasPage() {
             setConnections((prev) => [...prev, { id: nanoid(), ...connection }]);
             setSelectedNodeIds(new Set([newNode.id]));
             setSelectedConnectionId(null);
-            if (type !== CanvasNodeType.Text) setDialogNodeId(newNode.id);
+            if (type !== CanvasNodeType.Text && type !== CanvasNodeType.Audio) setDialogNodeId(newNode.id);
             setPendingConnectionCreate(null);
             setConnecting(null);
         },
@@ -611,7 +613,7 @@ function InfiniteCanvasPage() {
             setNodes((prev) => [...prev, newNode]);
             setSelectedNodeIds(new Set([newNode.id]));
             setSelectedConnectionId(null);
-            if (type !== CanvasNodeType.Text) setDialogNodeId(newNode.id);
+            if (type !== CanvasNodeType.Text && type !== CanvasNodeType.Audio) setDialogNodeId(newNode.id);
         },
         [effectiveConfig.imageModel, effectiveConfig.model, effectiveConfig.size, getCanvasCenter],
     );
@@ -1111,6 +1113,26 @@ function InfiniteCanvasPage() {
         setDialogNodeId(id);
     }, []);
 
+    const createAudioFileNode = useCallback(async (file: File, position: Position) => {
+        const audio = await uploadMediaFile(file, "audio");
+        const spec = NODE_DEFAULT_SIZE[CanvasNodeType.Audio];
+        const id = `audio-${Date.now()}-${Math.random().toString(36).slice(2, 7)}`;
+        setNodes((prev) => [
+            ...prev,
+            {
+                id,
+                type: CanvasNodeType.Audio,
+                title: file.name,
+                position: { x: position.x - spec.width / 2, y: position.y - spec.height / 2 },
+                width: spec.width,
+                height: spec.height,
+                metadata: audioMetadata(audio),
+            },
+        ]);
+        setSelectedNodeIds(new Set([id]));
+        setSelectedConnectionId(null);
+    }, []);
+
     const createTextNodeFromClipboard = useCallback(
         (text: string) => {
             const trimmed = text.trim();
@@ -1324,8 +1346,8 @@ function InfiniteCanvasPage() {
     }, []);
 
     const downloadNodeImage = useCallback((node: CanvasNodeData) => {
-        if ((node.type !== CanvasNodeType.Image && node.type !== CanvasNodeType.Video) || !node.metadata?.content) return;
-        saveAs(node.metadata.content, `canvas-${node.type}-${node.id}.${node.type === CanvasNodeType.Video ? "mp4" : imageExtension(node.metadata.content)}`);
+        if ((node.type !== CanvasNodeType.Image && node.type !== CanvasNodeType.Video && node.type !== CanvasNodeType.Audio) || !node.metadata?.content) return;
+        saveAs(node.metadata.content, `canvas-${node.type}-${node.id}.${node.type === CanvasNodeType.Video ? "mp4" : node.type === CanvasNodeType.Audio ? audioExtension(node.metadata.mimeType) : imageExtension(node.metadata.content)}`);
     }, []);
 
     const saveNodeAsset = useCallback(
@@ -1453,9 +1475,19 @@ function InfiniteCanvasPage() {
         async (event: ReactChangeEvent<HTMLInputElement>) => {
             const file = event.target.files?.[0];
             const target = uploadTargetRef.current;
-            if (!file || (!file.type.startsWith("image/") && !file.type.startsWith("video/"))) return;
+            if (!file || (!file.type.startsWith("image/") && !file.type.startsWith("video/") && !isAudioFile(file))) return;
 
             if (target?.nodeId) {
+                if (isAudioFile(file)) {
+                    const audio = await uploadMediaFile(file, "audio");
+                    const spec = NODE_DEFAULT_SIZE[CanvasNodeType.Audio];
+                    setNodes((prev) => prev.map((node) => (node.id === target.nodeId ? { ...node, type: CanvasNodeType.Audio, title: file.name, position: { x: node.position.x + node.width / 2 - spec.width / 2, y: node.position.y + node.height / 2 - spec.height / 2 }, width: spec.width, height: spec.height, metadata: { ...node.metadata, ...audioMetadata(audio), errorDetails: undefined } } : node)));
+                    setSelectedNodeIds(new Set([target.nodeId]));
+                    setSelectedConnectionId(null);
+                    uploadTargetRef.current = null;
+                    event.target.value = "";
+                    return;
+                }
                 if (file.type.startsWith("video/")) {
                     const video = await uploadMediaFile(file, "video");
                     const nextSize = fitNodeSize(video.width || 1280, video.height || 720, VIDEO_NODE_MAX_WIDTH, VIDEO_NODE_MAX_HEIGHT);
@@ -1505,25 +1537,25 @@ function InfiniteCanvasPage() {
                 setDialogNodeId(target.nodeId);
             } else {
                 const position = target?.position || screenToCanvas((containerRef.current?.getBoundingClientRect().left || 0) + size.width / 2, (containerRef.current?.getBoundingClientRect().top || 0) + size.height / 2);
-                void (file.type.startsWith("video/") ? createVideoFileNode(file, position) : createImageFileNode(file, position));
+                void (isAudioFile(file) ? createAudioFileNode(file, position) : file.type.startsWith("video/") ? createVideoFileNode(file, position) : createImageFileNode(file, position));
             }
 
             uploadTargetRef.current = null;
             event.target.value = "";
         },
-        [createImageFileNode, createVideoFileNode, screenToCanvas, size.height, size.width],
+        [createAudioFileNode, createImageFileNode, createVideoFileNode, screenToCanvas, size.height, size.width],
     );
 
     const handleDrop = useCallback(
         (event: ReactDragEvent<HTMLDivElement>) => {
             event.preventDefault();
-            const file = Array.from(event.dataTransfer.files).find((item) => item.type.startsWith("image/") || item.type.startsWith("video/"));
+            const file = Array.from(event.dataTransfer.files).find((item) => item.type.startsWith("image/") || item.type.startsWith("video/") || isAudioFile(item));
             if (!file) return;
 
             const pos = screenToCanvas(event.clientX, event.clientY);
-            void (file.type.startsWith("video/") ? createVideoFileNode(file, pos) : createImageFileNode(file, pos));
+            void (isAudioFile(file) ? createAudioFileNode(file, pos) : file.type.startsWith("video/") ? createVideoFileNode(file, pos) : createImageFileNode(file, pos));
         },
-        [createImageFileNode, createVideoFileNode, screenToCanvas],
+        [createAudioFileNode, createImageFileNode, createVideoFileNode, screenToCanvas],
     );
 
     const pasteAssistantImage = useCallback(
@@ -1749,14 +1781,14 @@ function InfiniteCanvasPage() {
                         position: isEmptyVideoNode ? sourceNode.position : { x: parent.x + (sourceNode?.width || spec.width) + 96, y: parent.y },
                         width: isEmptyVideoNode ? sourceNode.width : spec.width,
                         height: isEmptyVideoNode ? sourceNode.height : spec.height,
-                        metadata: { prompt: effectivePrompt, status: NODE_STATUS_LOADING, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, references: generationReferenceUrls(generationContext) },
+                        metadata: { prompt: effectivePrompt, status: NODE_STATUS_LOADING, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, generateAudio: generationConfig.videoGenerateAudio, watermark: generationConfig.videoWatermark, references: generationReferenceUrls(generationContext) },
                     };
                     pendingChildIds = [videoId];
                     setNodes((prev) => (isEmptyVideoNode ? prev.map((node) => (node.id === nodeId ? { ...node, ...videoNode } : node)) : [...prev.map((node) => (node.id === nodeId ? { ...node, metadata: { ...node.metadata, status: NODE_STATUS_SUCCESS } } : node)), videoNode]));
                     if (!isEmptyVideoNode) setConnections((prev) => [...prev, { id: nanoid(), fromNodeId: nodeId, toNodeId: videoId }]);
-                    const video = await storeGeneratedVideo(await requestVideoGeneration(generationConfig, effectivePrompt, generationContext.referenceImages, generationContext.referenceVideos));
+                    const video = await storeGeneratedVideo(await requestVideoGeneration(generationConfig, effectivePrompt, generationContext.referenceImages, generationContext.referenceVideos, generationContext.referenceAudios));
                     const videoSize = fitNodeSize(video.width || spec.width, video.height || spec.height, VIDEO_NODE_MAX_WIDTH, VIDEO_NODE_MAX_HEIGHT);
-                    setNodes((prev) => prev.map((node) => (node.id === videoId ? { ...node, width: videoSize.width, height: videoSize.height, position: { x: node.position.x + node.width / 2 - videoSize.width / 2, y: node.position.y + node.height / 2 - videoSize.height / 2 }, metadata: { ...node.metadata, ...videoMetadata(video), prompt: effectivePrompt, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, references: generationReferenceUrls(generationContext) } } : node)));
+                    setNodes((prev) => prev.map((node) => (node.id === videoId ? { ...node, width: videoSize.width, height: videoSize.height, position: { x: node.position.x + node.width / 2 - videoSize.width / 2, y: node.position.y + node.height / 2 - videoSize.height / 2 }, metadata: { ...node.metadata, ...videoMetadata(video), prompt: effectivePrompt, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, generateAudio: generationConfig.videoGenerateAudio, watermark: generationConfig.videoWatermark, references: generationReferenceUrls(generationContext) } } : node)));
                     return;
                 }
 
@@ -1874,9 +1906,9 @@ function InfiniteCanvasPage() {
                     return;
                 }
                 if (node.type === CanvasNodeType.Video) {
-                    const video = await storeGeneratedVideo(await requestVideoGeneration(generationConfig, prompt, retryImages, context?.referenceVideos || []));
+                    const video = await storeGeneratedVideo(await requestVideoGeneration(generationConfig, prompt, retryImages, context?.referenceVideos || [], context?.referenceAudios || []));
                     const videoSize = fitNodeSize(video.width || node.width, video.height || node.height, VIDEO_NODE_MAX_WIDTH, VIDEO_NODE_MAX_HEIGHT);
-                    setNodes((prev) => prev.map((item) => (item.id === node.id ? { ...item, width: videoSize.width, height: videoSize.height, position: { x: item.position.x + item.width / 2 - videoSize.width / 2, y: item.position.y + item.height / 2 - videoSize.height / 2 }, metadata: { ...item.metadata, ...videoMetadata(video), prompt, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality } } : item)));
+                    setNodes((prev) => prev.map((item) => (item.id === node.id ? { ...item, width: videoSize.width, height: videoSize.height, position: { x: item.position.x + item.width / 2 - videoSize.width / 2, y: item.position.y + item.height / 2 - videoSize.height / 2 }, metadata: { ...item.metadata, ...videoMetadata(video), prompt, model: generationConfig.model, size: generationConfig.size, seconds: generationConfig.videoSeconds, vquality: generationConfig.vquality, generateAudio: generationConfig.videoGenerateAudio, watermark: generationConfig.videoWatermark } } : item)));
                     return;
                 }
 
@@ -2196,6 +2228,7 @@ function InfiniteCanvasPage() {
                     showImageInfo={showImageInfo}
                     onAddImage={() => createNode(CanvasNodeType.Image)}
                     onAddVideo={() => createNode(CanvasNodeType.Video)}
+                    onAddAudio={() => createNode(CanvasNodeType.Audio)}
                     onAddText={() => createNode(CanvasNodeType.Text)}
                     onAddConfig={() => createNode(CanvasNodeType.Config)}
                     onUndo={undoCanvas}
@@ -2235,7 +2268,7 @@ function InfiniteCanvasPage() {
                     />
                 ) : null}
 
-                <input ref={imageInputRef} type="file" accept="image/*,video/*" className="hidden" onChange={handleImageInputChange} />
+                <input ref={imageInputRef} type="file" accept="image/*,video/*,audio/mpeg,audio/wav,audio/x-wav,.mp3,.wav" className="hidden" onChange={handleImageInputChange} />
 
                 <CanvasNodeInfoModal node={infoNode} open={Boolean(infoNode)} onClose={() => setInfoNodeId(null)} />
 
@@ -2377,7 +2410,7 @@ function CanvasTopBar({
                                 { key: "new", icon: <Plus className="size-4" />, label: "新建画布", onClick: onCreateProject },
                                 { key: "delete", danger: true, icon: <Trash2 className="size-4" />, label: "删除当前画布", onClick: onDeleteProject },
                                 { type: "divider" },
-                                { key: "import", icon: <Upload className="size-4" />, label: "导入图片", onClick: onImportImage },
+                                { key: "import", icon: <Upload className="size-4" />, label: "导入素材", onClick: onImportImage },
                                 { type: "divider" },
                                 { key: "undo", disabled: !canUndo, icon: <Undo2 className="size-4" />, label: <MenuLabel text="撤销" shortcut="⌘ Z" />, onClick: onUndo },
                                 { key: "redo", disabled: !canRedo, icon: <Redo2 className="size-4" />, label: <MenuLabel text="重做" shortcut="⌘ ⇧ Z / ⌘ Y" />, onClick: onRedo },
@@ -2458,7 +2491,7 @@ function CanvasTopBar({
                     <Shortcut keys={["Ctrl / Cmd", "Y"]} value="重做" />
                     <Shortcut keys={["Delete / Backspace"]} value="删除选中" />
                     <Shortcut keys={["Esc"]} value="取消选择并关闭浮层" />
-                    <Shortcut keys={["拖入图片"]} value="上传到画布" />
+                    <Shortcut keys={["拖入图片/视频/音频"]} value="上传到画布" />
                 </div>
             </Modal>
         </>
@@ -2499,12 +2532,20 @@ function imageExtension(dataUrl: string) {
     return dataUrl.match(/^data:image[/]([^;]+)/)?.[1] || dataUrl.match(/image[/]([^;]+)/)?.[1] || "png";
 }
 
+function audioExtension(mimeType?: string) {
+    return mimeType?.includes("wav") ? "wav" : "mp3";
+}
+
 function imageMetadata(image: UploadedImage): CanvasNodeMetadata {
     return { content: image.url, storageKey: image.storageKey, status: "success", naturalWidth: image.width, naturalHeight: image.height, bytes: image.bytes, mimeType: image.mimeType };
 }
 
 function videoMetadata(video: UploadedFile): CanvasNodeMetadata {
-    return { content: video.url, storageKey: video.storageKey, status: "success", naturalWidth: video.width, naturalHeight: video.height, bytes: video.bytes, mimeType: video.mimeType || "video/mp4" };
+    return { content: video.url, storageKey: video.storageKey, status: "success", naturalWidth: video.width, naturalHeight: video.height, bytes: video.bytes, mimeType: video.mimeType || "video/mp4", durationMs: video.durationMs };
+}
+
+function audioMetadata(audio: UploadedFile): CanvasNodeMetadata {
+    return { content: audio.url, storageKey: audio.storageKey, status: "success", bytes: audio.bytes, mimeType: audio.mimeType || "audio/mpeg", durationMs: audio.durationMs };
 }
 
 function buildImageGenerationMetadata(type: CanvasImageGenerationType, config: AiConfig, count: number, references: ReferenceImage[]): CanvasNodeMetadata {
@@ -2522,10 +2563,11 @@ function referenceUrl(image: ReferenceImage) {
     return image.storageKey || image.url || (!image.dataUrl.startsWith("data:") ? image.dataUrl : undefined);
 }
 
-function generationReferenceUrls(context: { referenceImages: ReferenceImage[]; referenceVideos: Array<{ storageKey?: string; url?: string }> }) {
+function generationReferenceUrls(context: { referenceImages: ReferenceImage[]; referenceVideos: Array<{ storageKey?: string; url?: string }>; referenceAudios?: Array<{ storageKey?: string; url?: string }> }) {
     return [
         ...context.referenceImages.map(referenceUrl).filter((url): url is string => Boolean(url)),
         ...context.referenceVideos.map((video) => video.storageKey || video.url).filter((url): url is string => Boolean(url)),
+        ...(context.referenceAudios || []).map((audio) => audio.storageKey || audio.url).filter((url): url is string => Boolean(url)),
     ];
 }
 
@@ -2545,7 +2587,7 @@ async function hydrateCanvasImages(nodes: CanvasNodeData[]) {
     return Promise.all(
         nodes.map(async (node) => {
             const content = node.metadata?.content;
-            if (node.type === CanvasNodeType.Video && node.metadata?.storageKey) return { ...node, metadata: { ...node.metadata, content: await resolveMediaUrl(node.metadata.storageKey, content) } };
+            if ((node.type === CanvasNodeType.Video || node.type === CanvasNodeType.Audio) && node.metadata?.storageKey) return { ...node, metadata: { ...node.metadata, content: await resolveMediaUrl(node.metadata.storageKey, content) } };
             if (node.type !== CanvasNodeType.Image || !content) return node;
             if (node.metadata?.storageKey) return { ...node, metadata: { ...node.metadata, content: await resolveImageUrl(node.metadata.storageKey, content) } };
             if (!content.startsWith("data:image/")) return node;
@@ -2605,6 +2647,7 @@ function getInputSummary(inputs: NodeGenerationInput[]) {
         textCount: inputs.filter((input) => input.type === "text").length,
         imageCount: inputs.filter((input) => input.type === "image").length,
         videoCount: inputs.filter((input) => input.type === "video").length,
+        audioCount: inputs.filter((input) => input.type === "audio").length,
     };
 }
 
@@ -2617,6 +2660,8 @@ function buildGenerationConfig(config: AiConfig, node: CanvasNodeData | undefine
         size: node?.metadata?.size || config.size || defaultConfig.size,
         videoSeconds: node?.metadata?.seconds || config.videoSeconds || defaultConfig.videoSeconds,
         vquality: node?.metadata?.vquality || config.vquality || defaultConfig.vquality,
+        videoGenerateAudio: node?.metadata?.generateAudio || config.videoGenerateAudio || defaultConfig.videoGenerateAudio,
+        videoWatermark: node?.metadata?.watermark || config.videoWatermark || defaultConfig.videoWatermark,
         count: String(node?.metadata?.count || (mode === "image" ? 3 : config.count) || defaultConfig.count),
     };
 }
@@ -2652,6 +2697,10 @@ function sourceNodeReferenceImages(node: CanvasNodeData | null) {
     ];
 }
 
+function isAudioFile(file: File) {
+    return file.type.startsWith("audio/") || /\.(mp3|wav)$/i.test(file.name);
+}
+
 function isHiddenBatchChild(node: CanvasNodeData, nodes: CanvasNodeData[], collapsingBatchIds?: Set<string>) {
     const rootId = node.metadata?.batchRootId;
     if (!rootId) return false;
diff --git a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
index 216b814..8966229 100644
--- a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
@@ -2,7 +2,7 @@
 
 import type { CSSProperties } from "react";
 import { useState } from "react";
-import { ArrowDown, ArrowLeft, ArrowRight, ArrowUp, Edit3, Eye, Image as ImageIcon, LoaderCircle, MessageSquare, Play, Video } from "lucide-react";
+import { ArrowDown, ArrowLeft, ArrowRight, ArrowUp, Edit3, Eye, Image as ImageIcon, LoaderCircle, MessageSquare, Music2, Play, Video } from "lucide-react";
 import { App, Button, Empty, Input, Modal, Segmented } from "antd";
 
 import { ModelPicker } from "@/components/model-picker";
@@ -18,7 +18,7 @@ import type { CanvasGenerationMode, CanvasNodeData, CanvasNodeMetadata } from ".
 type CanvasConfigNodePanelProps = {
     node: CanvasNodeData;
     isRunning: boolean;
-    inputSummary: { textCount: number; imageCount: number; videoCount: number };
+    inputSummary: { textCount: number; imageCount: number; videoCount: number; audioCount: number };
     inputs: NodeGenerationInput[];
     onConfigChange: (nodeId: string, patch: Partial<CanvasNodeMetadata>) => void;
     onTextInputChange: (nodeId: string, content: string) => void;
@@ -42,6 +42,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
     const textInputs = inputs.filter((input) => input.type === "text");
     const imageInputs = inputs.filter((input) => input.type === "image");
     const videoInputs = inputs.filter((input) => input.type === "video");
+    const audioInputs = inputs.filter((input) => input.type === "audio");
 
     const moveInput = (input: NodeGenerationInput, offset: number) => {
         const sameTypeInputs = inputs.filter((item) => item.type === input.type);
@@ -114,6 +115,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
                 <InputChip label="提示词" value={`${inputSummary.textCount} 个`} style={chipStyle} />
                 <InputChip label="参考图" value={`${inputSummary.imageCount} 张`} style={chipStyle} />
                 <InputChip label="参考视频" value={`${inputSummary.videoCount} 个`} style={chipStyle} />
+                <InputChip label="参考音频" value={`${inputSummary.audioCount} 个`} style={chipStyle} />
                 <button type="button" className="inline-flex h-7 cursor-pointer items-center gap-1 rounded-md border px-2 text-[11px]" style={chipStyle} onClick={() => setPreviewOpen(true)}>
                     <Eye className="size-3.5" />
                     预览
@@ -123,7 +125,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
             <div className={`mb-2 grid min-w-0 cursor-default items-center gap-2 ${mode === "text" ? "grid-cols-1" : "grid-cols-[minmax(0,1fr)_148px]"}`} onMouseDown={(event) => event.stopPropagation()}>
                 <ModelPicker className="canvas-compact-control h-10" config={config} value={config.model} onChange={(model) => onConfigChange(node.id, { model })} onMissingConfig={() => openConfigDialog(true)} fullWidth />
                 {mode === "video" ? (
-                    <CanvasVideoSettingsPopover config={config} placement="topRight" buttonClassName="canvas-compact-control !h-10 !w-full !justify-start !rounded-lg !px-2" onConfigChange={(key, value) => onConfigChange(node.id, key === "videoSeconds" ? { seconds: value } : { [key]: value })} />
+                    <CanvasVideoSettingsPopover config={config} placement="topRight" buttonClassName="canvas-compact-control !h-10 !w-full !justify-start !rounded-lg !px-2" onConfigChange={(key, value) => onConfigChange(node.id, videoConfigPatch(key, value))} />
                 ) : mode === "image" ? (
                     <CanvasImageSettingsPopover config={config} placement="topRight" autoAdjustOverflow={false} buttonClassName="canvas-compact-control !h-10 !w-full !justify-start !rounded-lg !px-2" onConfigChange={(key, value) => onConfigChange(node.id, key === "count" ? { count: Number(value) || 1 } : { [key]: value })} />
                 ) : null}
@@ -132,7 +134,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
             <Button
                 type="primary"
                 className="mt-auto !h-9 !w-full !cursor-pointer !rounded-lg"
-                disabled={isRunning || (!inputSummary.textCount && !inputSummary.imageCount && !inputSummary.videoCount)}
+                disabled={isRunning || (!inputSummary.textCount && !inputSummary.imageCount && !inputSummary.videoCount && !inputSummary.audioCount)}
                 onMouseDown={(event) => event.stopPropagation()}
                 onClick={() => onGenerate(node.id)}
             >
@@ -182,6 +184,15 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, inputs, o
                                     </div>
                                 </PreviewSection>
                             </div>
+                            <div className="shrink-0">
+                                <PreviewSection title="参考音频" count={audioInputs.length} empty="暂无参考音频">
+                                    <div className="thin-scrollbar flex gap-1.5 overflow-x-auto pb-1">
+                                        {audioInputs.map((input, index) => (
+                                            <AudioSortCard key={input.nodeId} input={input} audioIndex={index} audioTotal={audioInputs.length} theme={theme} onMove={moveInput} />
+                                        ))}
+                                    </div>
+                                </PreviewSection>
+                            </div>
                             <div className="grid min-h-0 flex-1 grid-cols-2 gap-3 overflow-hidden">
                                 <div className="thin-scrollbar min-h-0 overflow-y-auto pr-1.5">
                                     <PreviewSection title="文本提示词" count={textInputs.length} empty="暂无文本提示词">
@@ -324,6 +335,36 @@ function VideoSortCard({
     );
 }
 
+function AudioSortCard({
+    input,
+    audioIndex,
+    audioTotal,
+    theme,
+    onMove,
+}: {
+    input: NodeGenerationInput;
+    audioIndex: number;
+    audioTotal: number;
+    theme: (typeof canvasThemes)[keyof typeof canvasThemes];
+    onMove: (input: NodeGenerationInput, offset: number) => void;
+}) {
+    if (!input.audio) return null;
+    return (
+        <div className="w-48 shrink-0 rounded-lg border p-2" style={{ background: theme.node.fill, borderColor: theme.node.stroke }}>
+            <div className="mb-1.5 flex min-w-0 items-center gap-1.5 text-[11px] opacity-70">
+                <Music2 className="size-3.5 shrink-0" />
+                <span className="truncate">{input.title}</span>
+            </div>
+            <audio src={input.audio.url} controls className="h-8 w-full" preload="metadata" />
+            <div className="mt-1 flex justify-between">
+                <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !p-0" icon={<ArrowLeft className="size-3" />} disabled={audioIndex <= 0} onClick={() => onMove(input, -1)} />
+                <span className="text-[10px] opacity-45">{audioIndex + 1}</span>
+                <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !p-0" icon={<ArrowRight className="size-3" />} disabled={audioIndex >= audioTotal - 1} onClick={() => onMove(input, 1)} />
+            </div>
+        </div>
+    );
+}
+
 function VerticalOrderButtons({ index, total, onMove }: { index: number; total: number; onMove: (offset: number) => void }) {
     return (
         <>
@@ -360,6 +401,15 @@ function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: Can
         size: node.metadata?.size || globalConfig.size || defaultConfig.size,
         videoSeconds: node.metadata?.seconds || globalConfig.videoSeconds || defaultConfig.videoSeconds,
         vquality: node.metadata?.vquality || globalConfig.vquality || defaultConfig.vquality,
+        videoGenerateAudio: node.metadata?.generateAudio || globalConfig.videoGenerateAudio || defaultConfig.videoGenerateAudio,
+        videoWatermark: node.metadata?.watermark || globalConfig.videoWatermark || defaultConfig.videoWatermark,
         count: String(node.metadata?.count || (mode === "image" ? 3 : globalConfig.count) || defaultConfig.count),
     };
 }
+
+function videoConfigPatch(key: keyof AiConfig, value: string) {
+    if (key === "videoSeconds") return { seconds: value };
+    if (key === "videoGenerateAudio") return { generateAudio: value };
+    if (key === "videoWatermark") return { watermark: value };
+    return { [key]: value };
+}
diff --git a/web/src/app/(user)/canvas/components/canvas-mini-map.tsx b/web/src/app/(user)/canvas/components/canvas-mini-map.tsx
index 4557a87..1bf04f1 100644
--- a/web/src/app/(user)/canvas/components/canvas-mini-map.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-mini-map.tsx
@@ -115,7 +115,7 @@ export function Minimap({ nodes, viewport, viewportSize, onViewportChange }: { n
             >
                 {nodes.map((node) => {
                     const pos = toMinimap(node.position.x, node.position.y);
-                    const color = node.type === CanvasNodeType.Image ? "#10b981" : node.type === CanvasNodeType.Config ? "#60a5fa" : theme.node.muted;
+                    const color = node.type === CanvasNodeType.Image ? "#10b981" : node.type === CanvasNodeType.Video ? "#f97316" : node.type === CanvasNodeType.Audio ? "#a855f7" : node.type === CanvasNodeType.Config ? "#60a5fa" : theme.node.muted;
                     return (
                         <div
                             key={node.id}
diff --git a/web/src/app/(user)/canvas/components/canvas-node-generation.ts b/web/src/app/(user)/canvas/components/canvas-node-generation.ts
index 3b4c09a..3a3f1a4 100644
--- a/web/src/app/(user)/canvas/components/canvas-node-generation.ts
+++ b/web/src/app/(user)/canvas/components/canvas-node-generation.ts
@@ -1,24 +1,27 @@
 import type { ChatCompletionMessage } from "@/services/api/image";
 import type { ReferenceImage } from "@/types/image";
-import type { ReferenceVideo } from "@/types/media";
+import type { ReferenceAudio, ReferenceVideo } from "@/types/media";
 import { CanvasNodeType, type CanvasConnection, type CanvasNodeData } from "../types";
 
 export type NodeGenerationContext = {
     prompt: string;
     referenceImages: ReferenceImage[];
     referenceVideos: ReferenceVideo[];
+    referenceAudios: ReferenceAudio[];
     textCount: number;
     imageCount: number;
     videoCount: number;
+    audioCount: number;
 };
 
 export type NodeGenerationInput = {
     nodeId: string;
-    type: "text" | "image" | "video";
+    type: "text" | "image" | "video" | "audio";
     title: string;
     text?: string;
     image?: ReferenceImage;
     video?: ReferenceVideo;
+    audio?: ReferenceAudio;
 };
 
 export function buildNodeGenerationContext(nodeId: string, nodes: CanvasNodeData[], connections: CanvasConnection[], prompt: string): NodeGenerationContext {
@@ -29,14 +32,17 @@ export function buildNodeGenerationContext(nodeId: string, nodes: CanvasNodeData
         .join("\n\n");
     const referenceImages = inputs.map((input) => input.image).filter((image): image is ReferenceImage => Boolean(image));
     const referenceVideos = inputs.map((input) => input.video).filter((video): video is ReferenceVideo => Boolean(video));
+    const referenceAudios = inputs.map((input) => input.audio).filter((audio): audio is ReferenceAudio => Boolean(audio));
 
     return {
         prompt: upstreamText ? `${prompt}\n\n${upstreamText}` : prompt,
         referenceImages,
         referenceVideos,
+        referenceAudios,
         textCount: inputs.filter((input) => input.type === "text").length,
         imageCount: referenceImages.length,
         videoCount: referenceVideos.length,
+        audioCount: referenceAudios.length,
     };
 }
 
@@ -46,6 +52,8 @@ export function buildNodeGenerationInputs(nodeId: string, nodes: CanvasNodeData[
         if (image) return [{ nodeId: node.id, type: "image" as const, title: node.title, image }];
         const video = readReferenceVideo(node);
         if (video) return [{ nodeId: node.id, type: "video" as const, title: node.title, video }];
+        const audio = readReferenceAudio(node);
+        if (audio) return [{ nodeId: node.id, type: "audio" as const, title: node.title, audio }];
         const text = readNodeTextInput(node);
         if (text) return [{ nodeId: node.id, type: "text" as const, title: node.title, text }];
         return [];
@@ -97,6 +105,18 @@ function readReferenceVideo(node: CanvasNodeData): ReferenceVideo | null {
     };
 }
 
+function readReferenceAudio(node: CanvasNodeData): ReferenceAudio | null {
+    if (node.type !== CanvasNodeType.Audio || !node.metadata?.content) return null;
+    return {
+        id: node.id,
+        name: `${node.title || node.id}.mp3`,
+        type: node.metadata.mimeType || "audio/mpeg",
+        url: node.metadata.content,
+        storageKey: node.metadata.storageKey,
+        durationMs: node.metadata.durationMs,
+    };
+}
+
 function getOrderedUpstreamNodes(nodeId: string, nodes: CanvasNodeData[], connections: CanvasConnection[]) {
     const target = nodes.find((node) => node.id === nodeId);
     const upstreamNodes = connections
diff --git a/web/src/app/(user)/canvas/components/canvas-node-hover-toolbar.tsx b/web/src/app/(user)/canvas/components/canvas-node-hover-toolbar.tsx
index f43c6b0..1c0a786 100644
--- a/web/src/app/(user)/canvas/components/canvas-node-hover-toolbar.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-node-hover-toolbar.tsx
@@ -2,7 +2,7 @@
 
 import { useEffect, useMemo, useState, type ReactNode } from "react";
 import { Modal, Segmented, Tooltip } from "antd";
-import { Camera, Download, FolderPlus, Image as ImageIcon, Info, Lock, LockOpen, Maximize2, MessageSquare, Minus, Pencil, Plus, RefreshCw, Scissors, Settings2, Trash2, Upload, Video } from "lucide-react";
+import { Camera, Download, FolderPlus, Image as ImageIcon, Info, Lock, LockOpen, Maximize2, MessageSquare, Minus, Music2, Pencil, Plus, RefreshCw, Scissors, Settings2, Trash2, Upload, Video } from "lucide-react";
 
 import { canvasThemes } from "@/lib/canvas-theme";
 import { formatBytes, getDataUrlByteSize } from "@/lib/image-utils";
@@ -58,13 +58,15 @@ export function CanvasNodeHoverToolbar({
     const top = viewport.y + node.position.y * viewport.k - 14;
     const isImage = node.type === CanvasNodeType.Image;
     const isVideo = node.type === CanvasNodeType.Video;
+    const isAudio = node.type === CanvasNodeType.Audio;
     const hasImage = isImage && Boolean(node.metadata?.content);
     const hasVideo = isVideo && Boolean(node.metadata?.content);
+    const hasAudio = isAudio && Boolean(node.metadata?.content);
     const isText = node.type === CanvasNodeType.Text;
     const isConfig = node.type === CanvasNodeType.Config;
     const canOpenDialog = isText || hasImage || isVideo;
     const canRetry = node.metadata?.status === "error";
-    const hasSpecificTools = canRetry || isText || isImage || isVideo || isConfig;
+    const hasSpecificTools = canRetry || isText || isImage || isVideo || isAudio || isConfig;
 
     return (
         <div
@@ -80,7 +82,7 @@ export function CanvasNodeHoverToolbar({
             {hasSpecificTools ? <ToolbarDivider /> : null}
             {canRetry ? <ToolbarAction title="重新生成" label="重试" icon={<RefreshCw className="size-4" />} onClick={() => onRetry(node)} /> : null}
             {hasImage || hasVideo || isText ? <ToolbarAction title="加入我的素材" label="存素材" icon={<FolderPlus className="size-4" />} onClick={() => onSaveAsset(node)} /> : null}
-            {hasImage || hasVideo ? <IconAction title={hasVideo ? "下载视频" : "下载图片"} icon={<Download className="size-5" />} onClick={() => onDownload(node)} /> : null}
+            {hasImage || hasVideo || hasAudio ? <IconAction title={hasAudio ? "下载音频" : hasVideo ? "下载视频" : "下载图片"} icon={<Download className="size-5" />} onClick={() => onDownload(node)} /> : null}
             {canOpenDialog ? <ToolbarAction title="编辑" label="编辑" icon={<MessageSquare className="size-4" />} onClick={() => onToggleDialog(node)} /> : null}
             {isText ? <ToolbarAction title="编辑文本" label="编辑文字" icon={<Pencil className="size-4" />} onClick={() => onEditText(node)} /> : null}
             {isText ? <ToolbarAction title="用文本生图" label="生图" icon={<ImageIcon className="size-4" />} onClick={() => onGenerateImage(node)} /> : null}
@@ -89,6 +91,7 @@ export function CanvasNodeHoverToolbar({
             {isText ? <ToolbarAction title="增大字号" label="放大" icon={<Plus className="size-4" />} onClick={() => onIncreaseFont(node)} /> : null}
             {isImage ? <ToolbarAction title={hasImage ? "替换图片" : "上传图片"} label={hasImage ? "替换图片" : "上传图片"} icon={<Upload className="size-4" />} onClick={() => onUpload(node)} /> : null}
             {isVideo ? <ToolbarAction title={hasVideo ? "替换视频" : "上传视频"} label={hasVideo ? "替换视频" : "上传视频"} icon={<Video className="size-4" />} onClick={() => onUpload(node)} /> : null}
+            {isAudio ? <ToolbarAction title={hasAudio ? "替换音频" : "上传音频"} label={hasAudio ? "替换音频" : "上传音频"} icon={<Music2 className="size-4" />} onClick={() => onUpload(node)} /> : null}
             {hasImage ? (
                 <ToolbarAction
                     title={node.metadata?.freeResize ? "切换为等比缩放" : "切换为自由比例"}
@@ -151,7 +154,7 @@ export function CanvasNodeInfoModal({ node, open, onClose }: { node: CanvasNodeD
                     {view === "info" ? (
                         <div className="thin-scrollbar h-full space-y-3 overflow-auto pr-1">
                             <InfoRow label="ID" value={node.id} />
-                            <InfoRow label="类型" value={node.type === CanvasNodeType.Text ? "文本" : node.type === CanvasNodeType.Image ? "图片" : node.type === CanvasNodeType.Video ? "视频" : "生成配置"} />
+                            <InfoRow label="类型" value={node.type === CanvasNodeType.Text ? "文本" : node.type === CanvasNodeType.Image ? "图片" : node.type === CanvasNodeType.Video ? "视频" : node.type === CanvasNodeType.Audio ? "音频" : "生成配置"} />
                             <InfoRow label="尺寸" value={`${Math.round(node.width)} x ${Math.round(node.height)}`} />
                             <InfoRow label="位置" value={`${Math.round(node.position.x)}, ${Math.round(node.position.y)}`} />
                             <InfoRow label="状态" value={node.metadata?.status || "idle"} />
diff --git a/web/src/app/(user)/canvas/components/canvas-node-prompt-panel.tsx b/web/src/app/(user)/canvas/components/canvas-node-prompt-panel.tsx
index 677d58d..62db038 100644
--- a/web/src/app/(user)/canvas/components/canvas-node-prompt-panel.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-node-prompt-panel.tsx
@@ -93,7 +93,7 @@ export function CanvasNodePromptPanel({ node, isRunning, onPromptChange, onConfi
                     ) : mode === "video" ? (
                         <>
                             <ModelPicker config={config} value={config.model} onChange={(model) => onConfigChange(node.id, { model })} onMissingConfig={() => openConfigDialog(true)} />
-                            <CanvasVideoSettingsPopover config={config} buttonClassName="!h-10 !max-w-[170px] !justify-start !rounded-full !px-3" onConfigChange={(key, value) => onConfigChange(node.id, key === "videoSeconds" ? { seconds: value } : { [key]: value })} />
+                            <CanvasVideoSettingsPopover config={config} buttonClassName="!h-10 !max-w-[170px] !justify-start !rounded-full !px-3" onConfigChange={(key, value) => onConfigChange(node.id, videoConfigPatch(key, value))} />
                         </>
                     ) : (
                         <ModelPicker config={config} value={config.model} onChange={(model) => onConfigChange(node.id, { model })} onMissingConfig={() => openConfigDialog(true)} />
@@ -132,6 +132,15 @@ function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: Can
         size: node.metadata?.size || globalConfig.size || defaultConfig.size,
         videoSeconds: node.metadata?.seconds || globalConfig.videoSeconds || defaultConfig.videoSeconds,
         vquality: node.metadata?.vquality || globalConfig.vquality || defaultConfig.vquality,
+        videoGenerateAudio: node.metadata?.generateAudio || globalConfig.videoGenerateAudio || defaultConfig.videoGenerateAudio,
+        videoWatermark: node.metadata?.watermark || globalConfig.videoWatermark || defaultConfig.videoWatermark,
         count: String(node.metadata?.count || (mode === "image" ? 3 : globalConfig.count) || defaultConfig.count),
     };
 }
+
+function videoConfigPatch(key: keyof AiConfig, value: string) {
+    if (key === "videoSeconds") return { seconds: value };
+    if (key === "videoGenerateAudio") return { generateAudio: value };
+    if (key === "videoWatermark") return { watermark: value };
+    return { [key]: value };
+}
diff --git a/web/src/app/(user)/canvas/components/canvas-node.tsx b/web/src/app/(user)/canvas/components/canvas-node.tsx
index 7074825..07bd043 100644
--- a/web/src/app/(user)/canvas/components/canvas-node.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-node.tsx
@@ -2,7 +2,7 @@
 
 import React, { useCallback, useEffect, useRef, useState } from "react";
 import type { ReactNode } from "react";
-import { ChevronRight, Image as ImageIcon, RefreshCw, Star, Video } from "lucide-react";
+import { ChevronRight, Image as ImageIcon, Music2, RefreshCw, Star, Video } from "lucide-react";
 
 import { canvasThemes } from "@/lib/canvas-theme";
 import { formatBytes } from "@/lib/image-utils";
@@ -99,6 +99,7 @@ export const CanvasNode = React.memo(function CanvasNode({
     const [isEditingContent, setIsEditingContent] = useState(false);
     const hasImageContent = data.type === CanvasNodeType.Image && Boolean(data.metadata?.content);
     const hasVideoContent = data.type === CanvasNodeType.Video && Boolean(data.metadata?.content);
+    const hasAudioContent = data.type === CanvasNodeType.Audio && Boolean(data.metadata?.content);
     const isBatchRoot = data.type === CanvasNodeType.Image && Boolean(data.metadata?.isBatchRoot) && batchCount > 1;
     const isBatchChild = data.type === CanvasNodeType.Image && Boolean(data.metadata?.batchRootId);
     const isActive = isConnectionTarget || isSelected || isFocusRelated;
@@ -301,7 +302,7 @@ export const CanvasNode = React.memo(function CanvasNode({
 
                 {showImageInfo && hasImageContent ? <ImageInfoBar node={data} /> : null}
 
-                {!hasImageContent && !hasVideoContent ? <div className="pointer-events-none absolute inset-x-0 bottom-0 h-12" style={{ background: `linear-gradient(to top, ${theme.canvas.background}66, transparent)` }} /> : null}
+                {!hasImageContent && !hasVideoContent && !hasAudioContent ? <div className="pointer-events-none absolute inset-x-0 bottom-0 h-12" style={{ background: `linear-gradient(to top, ${theme.canvas.background}66, transparent)` }} /> : null}
 
                 <ResizeHandle corner="top-left" onMouseDown={handleResizeMouseDown} />
                 <ResizeHandle corner="top-right" onMouseDown={handleResizeMouseDown} />
@@ -332,6 +333,7 @@ const nodeContentRenderers = {
     [CanvasNodeType.Image]: ImageNodeContent,
     [CanvasNodeType.Config]: EmptyImageContent,
     [CanvasNodeType.Video]: VideoNodeContent,
+    [CanvasNodeType.Audio]: AudioNodeContent,
 } satisfies Record<CanvasNodeType, (props: NodeContentRendererProps) => ReactNode>;
 
 function LoadingContent({ theme }: Pick<NodeContentRendererProps, "theme">) {
@@ -472,6 +474,25 @@ function VideoNodeContent({ node, theme }: NodeContentRendererProps) {
     return <video src={node.metadata.content} controls className="h-full w-full rounded-[18px] bg-black object-contain" data-canvas-no-zoom />;
 }
 
+function AudioNodeContent({ node, theme }: NodeContentRendererProps) {
+    if (!node.metadata?.content)
+        return (
+            <div className="flex h-full w-full flex-col items-center justify-center gap-2" style={{ color: theme.node.placeholder }}>
+                <Music2 className="size-7 opacity-35" />
+                <span className="text-sm">空音频节点</span>
+            </div>
+        );
+    return (
+        <div className="flex h-full w-full flex-col justify-center gap-3 px-4" style={{ background: theme.node.fill, color: theme.node.text }}>
+            <div className="flex min-w-0 items-center gap-2 text-sm opacity-70">
+                <Music2 className="size-4 shrink-0" />
+                <span className="truncate">{node.title || "音频"}</span>
+            </div>
+            <audio src={node.metadata.content} controls className="w-full" data-canvas-no-zoom />
+        </div>
+    );
+}
+
 function ImageContent({
     node,
     isBatchRoot,
diff --git a/web/src/app/(user)/canvas/components/canvas-toolbar.tsx b/web/src/app/(user)/canvas/components/canvas-toolbar.tsx
index fe229d8..e7eca92 100644
--- a/web/src/app/(user)/canvas/components/canvas-toolbar.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-toolbar.tsx
@@ -1,7 +1,7 @@
 import type { CSSProperties, MouseEvent as ReactMouseEvent, ReactNode, RefObject } from "react";
 import { useRef, useState } from "react";
 import { Button, Segmented, Switch } from "antd";
-import { CircleDot, Eraser, FolderOpen, Grid2x2, Hand, Image as ImageIcon, Info, Library, Moon, Palette, Redo2, Settings2, Square, Sun, Trash2, Type, Undo2, Upload, Video } from "lucide-react";
+import { CircleDot, Eraser, FolderOpen, Grid2x2, Hand, Image as ImageIcon, Info, Library, Moon, Music2, Palette, Redo2, Settings2, Square, Sun, Trash2, Type, Undo2, Upload, Video } from "lucide-react";
 
 import { canvasThemes, type CanvasBackgroundMode, type CanvasColorTheme, type CanvasTheme } from "@/lib/canvas-theme";
 import { useThemeStore } from "@/stores/use-theme-store";
@@ -15,6 +15,7 @@ export function CanvasToolbar({
     showImageInfo,
     onAddImage,
     onAddVideo,
+    onAddAudio,
     onAddText,
     onAddConfig,
     onUndo,
@@ -35,6 +36,7 @@ export function CanvasToolbar({
     showImageInfo: boolean;
     onAddImage: () => void;
     onAddVideo: () => void;
+    onAddAudio: () => void;
     onAddText: () => void;
     onAddConfig: () => void;
     onUndo: () => void;
@@ -84,10 +86,13 @@ export function CanvasToolbar({
                 <ToolbarButton id="tool-video" label="视频" hovered={hovered} hoverStyle={hoverStyle} wrapRef={wrapRef} onTipX={setTipX} onHover={setHovered} onClick={onAddVideo}>
                     <Video className="size-4.5" />
                 </ToolbarButton>
+                <ToolbarButton id="tool-audio" label="音频" hovered={hovered} hoverStyle={hoverStyle} wrapRef={wrapRef} onTipX={setTipX} onHover={setHovered} onClick={onAddAudio}>
+                    <Music2 className="size-4.5" />
+                </ToolbarButton>
                 <ToolbarButton id="tool-config" label="生成配置" hovered={hovered} hoverStyle={hoverStyle} wrapRef={wrapRef} onTipX={setTipX} onHover={setHovered} onClick={onAddConfig}>
                     <Settings2 className="size-4.5" />
                 </ToolbarButton>
-                <ToolbarButton id="tool-upload" label="上传图片" hovered={hovered} hoverStyle={hoverStyle} wrapRef={wrapRef} onTipX={setTipX} onHover={setHovered} onClick={onUpload}>
+                <ToolbarButton id="tool-upload" label="上传素材" hovered={hovered} hoverStyle={hoverStyle} wrapRef={wrapRef} onTipX={setTipX} onHover={setHovered} onClick={onUpload}>
                     <Upload className="size-4.5" />
                 </ToolbarButton>
                 <Divider theme={theme} />
@@ -279,8 +284,9 @@ function toolLabel(id: string) {
     if (id === "tool-text") return "文本";
     if (id === "tool-image") return "图片";
     if (id === "tool-video") return "视频";
+    if (id === "tool-audio") return "音频";
     if (id === "tool-config") return "生成配置";
-    if (id === "tool-upload") return "上传图片";
+    if (id === "tool-upload") return "上传素材";
     if (id === "tool-library") return "素材库";
     if (id === "tool-assets") return "我的素材";
     if (id === "tool-style") return "画布外观";
diff --git a/web/src/app/(user)/canvas/constants.ts b/web/src/app/(user)/canvas/constants.ts
index 57e2925..3029ec7 100644
--- a/web/src/app/(user)/canvas/constants.ts
+++ b/web/src/app/(user)/canvas/constants.ts
@@ -13,6 +13,7 @@ export const NODE_DEFAULT_SIZE = {
     [CanvasNodeType.Text]: { width: 340, height: 240, title: "Note" },
     [CanvasNodeType.Config]: { width: 340, height: 240, title: "生成配置" },
     [CanvasNodeType.Video]: { width: 420, height: 236, title: "Video" },
+    [CanvasNodeType.Audio]: { width: 340, height: 120, title: "Audio" },
 } satisfies Record<CanvasNodeType, { width: number; height: number; title: string }>;
 
 export const NODE_SPECS = {
@@ -32,6 +33,10 @@ export const NODE_SPECS = {
         ...NODE_DEFAULT_SIZE[CanvasNodeType.Video],
         metadata: { content: "", status: "idle" },
     },
+    [CanvasNodeType.Audio]: {
+        ...NODE_DEFAULT_SIZE[CanvasNodeType.Audio],
+        metadata: { content: "", status: "idle" },
+    },
 } satisfies Record<CanvasNodeType, CanvasNodeSpec>;
 
 export function getNodeSpec(type: CanvasNodeType) {
diff --git a/web/src/app/(user)/canvas/types.ts b/web/src/app/(user)/canvas/types.ts
index 7321195..023b5f2 100644
--- a/web/src/app/(user)/canvas/types.ts
+++ b/web/src/app/(user)/canvas/types.ts
@@ -14,6 +14,7 @@ export enum CanvasNodeType {
     Text = "text",
     Config = "config",
     Video = "video",
+    Audio = "audio",
 }
 
 export type CanvasNodeStatus = "idle" | "success" | "loading" | "error";
@@ -34,6 +35,8 @@ export type CanvasNodeMetadata = {
     count?: number;
     seconds?: string;
     vquality?: string;
+    generateAudio?: string;
+    watermark?: string;
     references?: string[];
     naturalWidth?: number;
     naturalHeight?: number;
@@ -48,6 +51,7 @@ export type CanvasNodeMetadata = {
     storageKey?: string;
     mimeType?: string;
     bytes?: number;
+    durationMs?: number;
 };
 
 export type CanvasNodeData = {
diff --git a/web/src/app/(user)/video/page.tsx b/web/src/app/(user)/video/page.tsx
index 9205d70..7a6b009 100644
--- a/web/src/app/(user)/video/page.tsx
+++ b/web/src/app/(user)/video/page.tsx
@@ -1,6 +1,6 @@
 "use client";
 
-import { BookOpen, CheckSquare, ClipboardPaste, Download, FolderPlus, History, LoaderCircle, Plus, SlidersHorizontal, Sparkles, Trash2, Upload, VideoIcon } from "lucide-react";
+import { BookOpen, CheckSquare, ClipboardPaste, Download, FolderPlus, History, LoaderCircle, Music2, Plus, SlidersHorizontal, Sparkles, Trash2, Upload, VideoIcon } from "lucide-react";
 import { useEffect, useRef, useState } from "react";
 import { App, Button, Checkbox, Drawer, Empty, Input, Modal, Tag, Typography } from "antd";
 import localforage from "localforage";
@@ -13,6 +13,7 @@ import { PromptSelectDialog } from "@/components/prompts/prompt-select-dialog";
 import { VideoSettingsPanel, normalizeVideoResolutionValue, normalizeVideoSizeValue, videoSizeLabel } from "@/components/video-settings-panel";
 import { canvasThemes } from "@/lib/canvas-theme";
 import { formatBytes, formatDuration } from "@/lib/image-utils";
+import { boolConfig, isSeedanceVideoConfig, normalizeSeedanceRatio, SEEDANCE_REFERENCE_LIMITS } from "@/lib/seedance-video";
 import { deleteStoredMedia, resolveMediaUrl, uploadMediaFile } from "@/services/file-storage";
 import { resolveImageUrl, uploadImage } from "@/services/image-storage";
 import { requestVideoGeneration, storeGeneratedVideo } from "@/services/api/video";
@@ -20,7 +21,7 @@ import { useAssetStore } from "@/stores/use-asset-store";
 import { useConfigStore, useEffectiveConfig, type AiConfig } from "@/stores/use-config-store";
 import { useThemeStore } from "@/stores/use-theme-store";
 import type { ReferenceImage } from "@/types/image";
-import type { ReferenceVideo } from "@/types/media";
+import type { ReferenceAudio, ReferenceVideo } from "@/types/media";
 
 type GeneratedVideo = {
     id: string;
@@ -50,6 +51,7 @@ type GenerationLog = {
     config: GenerationLogConfig;
     references: ReferenceImage[];
     videoReferences: ReferenceVideo[];
+    audioReferences: ReferenceAudio[];
     durationMs: number;
     size: string;
     resolution: string;
@@ -59,7 +61,7 @@ type GenerationLog = {
     error?: string;
 };
 
-type GenerationLogConfig = Pick<AiConfig, "model" | "videoModel" | "size" | "vquality" | "videoSeconds">;
+type GenerationLogConfig = Pick<AiConfig, "model" | "videoModel" | "size" | "vquality" | "videoSeconds" | "videoGenerateAudio" | "videoWatermark">;
 
 type UpdateAiConfig = <K extends keyof AiConfig>(key: K, value: AiConfig[K]) => void;
 
@@ -78,6 +80,7 @@ export default function VideoPage() {
     const [prompt, setPrompt] = useState("");
     const [references, setReferences] = useState<ReferenceImage[]>([]);
     const [videoReferences, setVideoReferences] = useState<ReferenceVideo[]>([]);
+    const [audioReferences, setAudioReferences] = useState<ReferenceAudio[]>([]);
     const [results, setResults] = useState<GenerationResult[]>([]);
     const [logs, setLogs] = useState<GenerationLog[]>([]);
     const [running, setRunning] = useState(false);
@@ -106,8 +109,14 @@ export default function VideoPage() {
 
     const addReferences = async (files?: FileList | null) => {
         const selectedFiles = Array.from(files || []);
-        const imageFiles = selectedFiles.filter((file) => file.type.startsWith("image/")).slice(0, 7 - references.length);
-        const videoFiles = selectedFiles.filter((file) => file.type.startsWith("video/")).slice(0, 3 - videoReferences.length);
+        const unsupported = selectedFiles.filter((file) => !file.type.startsWith("image/") && !file.type.startsWith("video/") && !isSupportedAudioFile(file));
+        if (unsupported.length) message.warning("已忽略不支持的参考素材，请使用图片、mp4/mov 视频或 mp3/wav 音频");
+        const imageFiles = selectedFiles.filter((file) => file.type.startsWith("image/") && file.size <= SEEDANCE_REFERENCE_LIMITS.imageMaxBytes).slice(0, SEEDANCE_REFERENCE_LIMITS.images - references.length);
+        const videoFiles = selectedFiles.filter((file) => file.type.startsWith("video/") && file.size <= SEEDANCE_REFERENCE_LIMITS.videoMaxBytes).slice(0, SEEDANCE_REFERENCE_LIMITS.videos - videoReferences.length);
+        const audioFiles = selectedFiles.filter((file) => isSupportedAudioFile(file) && file.size <= SEEDANCE_REFERENCE_LIMITS.audioMaxBytes).slice(0, SEEDANCE_REFERENCE_LIMITS.audios - audioReferences.length);
+        if (selectedFiles.some((file) => file.type.startsWith("image/") && file.size > SEEDANCE_REFERENCE_LIMITS.imageMaxBytes)) message.warning("已忽略超过 30MB 的参考图");
+        if (selectedFiles.some((file) => file.type.startsWith("video/") && file.size > SEEDANCE_REFERENCE_LIMITS.videoMaxBytes)) message.warning("已忽略超过 50MB 的参考视频");
+        if (selectedFiles.some((file) => isSupportedAudioFile(file) && file.size > SEEDANCE_REFERENCE_LIMITS.audioMaxBytes)) message.warning("已忽略超过 15MB 的参考音频");
         const nextReferences = await Promise.all(
             imageFiles.map(async (file) => {
                 const image = await uploadImage(file);
@@ -117,11 +126,22 @@ export default function VideoPage() {
         const nextVideoReferences = await Promise.all(
             videoFiles.map(async (file) => {
                 const video = await uploadMediaFile(file, "video-reference");
-                return { id: nanoid(), name: file.name, type: video.mimeType, url: video.url, storageKey: video.storageKey };
+                return { id: nanoid(), name: file.name, type: video.mimeType, url: video.url, storageKey: video.storageKey, durationMs: video.durationMs };
             }),
         );
-        setReferences((value) => [...value, ...nextReferences].slice(0, 7));
-        setVideoReferences((value) => [...value, ...nextVideoReferences].slice(0, 3));
+        const nextAudioReferences = filterAudioReferencesByDuration(
+            audioReferences,
+            await Promise.all(
+                audioFiles.map(async (file) => {
+                    const audio = await uploadMediaFile(file, "audio-reference");
+                    return { id: nanoid(), name: file.name, type: audio.mimeType, url: audio.url, storageKey: audio.storageKey, durationMs: audio.durationMs };
+                }),
+            ),
+            message.warning,
+        );
+        setReferences((value) => [...value, ...nextReferences].slice(0, SEEDANCE_REFERENCE_LIMITS.images));
+        setVideoReferences((value) => [...value, ...nextVideoReferences].slice(0, SEEDANCE_REFERENCE_LIMITS.videos));
+        setAudioReferences((value) => [...value, ...nextAudioReferences].slice(0, SEEDANCE_REFERENCE_LIMITS.audios));
     };
 
     const addReferencesFromClipboard = async () => {
@@ -133,18 +153,17 @@ export default function VideoPage() {
                 return;
             }
             const nextReferences = await Promise.all(
-                blobs.slice(0, 7 - references.length).map(async (blob, index) => {
+                blobs.slice(0, SEEDANCE_REFERENCE_LIMITS.images - references.length).map(async (blob, index) => {
                     const image = await uploadImage(blob);
                     return { id: nanoid(), name: `clipboard-${index + 1}.png`, type: image.mimeType, dataUrl: image.url, storageKey: image.storageKey };
                 }),
             );
-            setReferences((value) => [...value, ...nextReferences].slice(0, 7));
+            setReferences((value) => [...value, ...nextReferences].slice(0, SEEDANCE_REFERENCE_LIMITS.images));
             message.success(`已读取 ${nextReferences.length} 张参考图`);
         } catch {
             message.error("剪切板里没有可读取的图片");
         }
     };
-
     const generate = async () => {
         const snapshot = buildRequestSnapshot();
         if (!snapshot) return;
@@ -155,7 +174,7 @@ export default function VideoPage() {
         const batchStartedAt = performance.now();
         setStartedAt(batchStartedAt);
         try {
-            const stored = await storeGeneratedVideo(await requestVideoGeneration(snapshot.config, snapshot.text, snapshot.references, snapshot.videoReferences));
+            const stored = await storeGeneratedVideo(await requestVideoGeneration(snapshot.config, snapshot.text, snapshot.references, snapshot.videoReferences, snapshot.audioReferences));
             const nextVideo: GeneratedVideo = {
                 id: nanoid(),
                 url: stored.url,
@@ -167,12 +186,12 @@ export default function VideoPage() {
                 mimeType: stored.mimeType,
             };
             setResults([{ id: nextVideo.id, status: "success", video: nextVideo }]);
-            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, videoReferences: snapshot.videoReferences, durationMs: nextVideo.durationMs, status: "成功", video: nextVideo }));
+            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, videoReferences: snapshot.videoReferences, audioReferences: snapshot.audioReferences, durationMs: nextVideo.durationMs, status: "成功", video: nextVideo }));
             message.success("视频已生成");
         } catch (error) {
             const errorMessage = error instanceof Error ? error.message : "生成失败";
             setResults([{ id: nanoid(), status: "failed", error: errorMessage }]);
-            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, videoReferences: snapshot.videoReferences, durationMs: performance.now() - batchStartedAt, status: "失败", error: errorMessage }));
+            saveLog(buildLog({ prompt: snapshot.text, model, config: snapshot.config, references: snapshot.references, videoReferences: snapshot.videoReferences, audioReferences: snapshot.audioReferences, durationMs: performance.now() - batchStartedAt, status: "失败", error: errorMessage }));
             message.error(errorMessage);
         } finally {
             setRunning(false);
@@ -190,7 +209,7 @@ export default function VideoPage() {
             openConfigDialog(true);
             return null;
         }
-        return { text, config: buildVideoConfig(effectiveConfig, model), references: [...references], videoReferences: [...videoReferences] };
+        return { text, config: buildVideoConfig(effectiveConfig, model), references: [...references], videoReferences: [...videoReferences], audioReferences: [...audioReferences] };
     };
 
     const retryResult = () => {
@@ -219,9 +238,9 @@ export default function VideoPage() {
             setPrompt(payload.content);
         } else if (payload.kind === "image") {
             const stored = await uploadImage(payload.dataUrl);
-            setReferences((value) => [...value, { id: nanoid(), name: payload.title, type: stored.mimeType, dataUrl: stored.url, storageKey: stored.storageKey }].slice(0, 7));
+            setReferences((value) => [...value, { id: nanoid(), name: payload.title, type: stored.mimeType, dataUrl: stored.url, storageKey: stored.storageKey }].slice(0, SEEDANCE_REFERENCE_LIMITS.images));
         } else if (payload.kind === "video") {
-            setVideoReferences((value) => [...value, { id: nanoid(), name: payload.title, type: "video/mp4", url: payload.url, storageKey: payload.storageKey }].slice(0, 3));
+            setVideoReferences((value) => [...value, { id: nanoid(), name: payload.title, type: "video/mp4", url: payload.url, storageKey: payload.storageKey }].slice(0, SEEDANCE_REFERENCE_LIMITS.videos));
         }
         setAssetPickerOpen(false);
     };
@@ -230,6 +249,7 @@ export default function VideoPage() {
         setPrompt("");
         setReferences([]);
         setVideoReferences([]);
+        setAudioReferences([]);
         setResults([]);
         setElapsedMs(0);
         setStartedAt(0);
@@ -263,10 +283,13 @@ export default function VideoPage() {
         setPrompt(log.prompt);
         setReferences(log.references || []);
         setVideoReferences(log.videoReferences || []);
+        setAudioReferences(log.audioReferences || []);
         if (log.config.videoModel || log.model) updateConfig("videoModel", log.config.videoModel || log.model);
         if (log.config.size) updateConfig("size", log.config.size);
         if (log.config.vquality) updateConfig("vquality", log.config.vquality);
         if (log.config.videoSeconds) updateConfig("videoSeconds", log.config.videoSeconds);
+        if (log.config.videoGenerateAudio) updateConfig("videoGenerateAudio", log.config.videoGenerateAudio);
+        if (log.config.videoWatermark) updateConfig("videoWatermark", log.config.videoWatermark);
         setResults(log.video ? [{ id: log.video.id, status: "success", video: log.video }] : [{ id: log.id, status: "failed", error: log.error || "生成失败" }]);
     };
 
@@ -328,7 +351,7 @@ export default function VideoPage() {
                                             </button>
                                         </div>
                                     ))}
-                                    {!references.length ? <div className="flex min-w-full items-center justify-center text-sm text-stone-500">暂无参考图，最多 7 张</div> : null}
+                                    {!references.length ? <div className="flex min-w-full items-center justify-center text-sm text-stone-500">暂无参考图，最多 9 张</div> : null}
                                 </div>
                             </div>
 
@@ -352,6 +375,30 @@ export default function VideoPage() {
                                 </div>
                             </div>
 
+                            <div className="min-w-0">
+                                <div className="mb-2 flex items-center justify-between gap-3">
+                                    <span className="text-base font-semibold">参考音频</span>
+                                    <Button size="small" icon={<Upload className="size-3.5" />} onClick={() => fileInputRef.current?.click()}>
+                                        上传
+                                    </Button>
+                                </div>
+                                <div className="hover-scrollbar hover-scrollbar-hint flex min-h-24 w-full min-w-0 max-w-full gap-2 overflow-x-scroll overflow-y-hidden rounded-lg border border-dashed border-stone-300 p-2 pb-3 overscroll-x-contain dark:border-stone-700">
+                                    {audioReferences.map((item) => (
+                                        <div key={item.id} className="group relative flex h-20 w-48 shrink-0 flex-col justify-center gap-2 rounded-md border border-stone-200 bg-stone-50 px-2 dark:border-stone-800 dark:bg-stone-900">
+                                            <div className="flex min-w-0 items-center gap-2 text-xs text-stone-500 dark:text-stone-400">
+                                                <Music2 className="size-4 shrink-0" />
+                                                <span className="truncate">{item.name}</span>
+                                            </div>
+                                            <audio src={item.url} controls className="h-8 w-full" preload="metadata" />
+                                            <button type="button" className="absolute right-1 top-1 hidden size-6 items-center justify-center rounded bg-black/60 text-white group-hover:flex" onClick={() => setAudioReferences((value) => value.filter((ref) => ref.id !== item.id))} aria-label="移除参考音频">
+                                                <Trash2 className="size-3.5" />
+                                            </button>
+                                        </div>
+                                    ))}
+                                    {!audioReferences.length ? <div className="flex min-w-full items-center justify-center text-center text-sm text-stone-500">暂无参考音频，最多 3 个，mp3/wav，单个 15MB 内</div> : null}
+                                </div>
+                            </div>
+
                             <div className="flex items-center justify-between rounded-lg border border-stone-200 bg-stone-50 px-3 py-2 text-sm dark:border-stone-800 dark:bg-stone-900 sm:hidden">
                                 <span className="truncate text-stone-500 dark:text-stone-400">
                                     {model} · {normalizeResolution(effectiveConfig.vquality)}p · {videoSizeLabel(effectiveConfig.size)} · {normalizeVideoSeconds(effectiveConfig.videoSeconds)}s
@@ -394,7 +441,7 @@ export default function VideoPage() {
             <input
                 ref={fileInputRef}
                 type="file"
-                accept="image/*,video/mp4,video/quicktime"
+                accept="image/*,video/mp4,video/quicktime,audio/mpeg,audio/wav,audio/x-wav,.mp3,.wav"
                 multiple
                 className="hidden"
                 onChange={(event) => {
@@ -583,6 +630,12 @@ async function normalizeLog(log: Partial<GenerationLog>): Promise<GenerationLog>
             url: item.storageKey ? await resolveMediaUrl(item.storageKey, item.url) : item.url,
         })),
     );
+    const audioReferences = await Promise.all(
+        (log.audioReferences || []).map(async (item) => ({
+            ...item,
+            url: item.storageKey ? await resolveMediaUrl(item.storageKey, item.url) : item.url,
+        })),
+    );
     const references = await Promise.all(
         (log.references || []).map(async (item) => ({
             ...item,
@@ -600,6 +653,7 @@ async function normalizeLog(log: Partial<GenerationLog>): Promise<GenerationLog>
         config,
         references,
         videoReferences,
+        audioReferences,
         durationMs: log.durationMs || 0,
         size: log.size || config.size || "",
         resolution: normalizeResolution(log.resolution || config.vquality || ""),
@@ -615,10 +669,35 @@ function serializeLog(log: GenerationLog): GenerationLog {
         ...log,
         references: log.references.map((item) => ({ ...item, dataUrl: item.storageKey ? "" : item.dataUrl })),
         videoReferences: log.videoReferences.map((item) => (item.storageKey ? { ...item, url: "" } : item)),
+        audioReferences: log.audioReferences.map((item) => (item.storageKey ? { ...item, url: "" } : item)),
         video: log.video?.storageKey ? { ...log.video, url: "" } : log.video,
     };
 }
 
+function isSupportedAudioFile(file: File) {
+    return file.type === "audio/mpeg" || file.type === "audio/mp3" || file.type === "audio/wav" || file.type === "audio/x-wav" || /\.(mp3|wav)$/i.test(file.name);
+}
+
+function filterAudioReferencesByDuration(existing: ReferenceAudio[], next: ReferenceAudio[], warn: (content: string) => void) {
+    let total = existing.reduce((sum, item) => sum + (item.durationMs || 0), 0);
+    const accepted: ReferenceAudio[] = [];
+    let skipped = false;
+    for (const item of next) {
+        if (item.durationMs && (item.durationMs < 2000 || item.durationMs > 15000)) {
+            skipped = true;
+            continue;
+        }
+        if (item.durationMs && total + item.durationMs > 15000) {
+            skipped = true;
+            continue;
+        }
+        total += item.durationMs || 0;
+        accepted.push(item);
+    }
+    if (skipped) warn("已忽略不符合时长要求的参考音频：单个 2-15 秒，总时长不超过 15 秒");
+    return accepted;
+}
+
 function normalizeLogConfig(log: Partial<GenerationLog>): GenerationLogConfig {
     return {
         model: log.config?.model || log.model || "",
@@ -626,16 +705,20 @@ function normalizeLogConfig(log: Partial<GenerationLog>): GenerationLogConfig {
         size: log.config?.size || log.size || "",
         vquality: normalizeResolution(log.config?.vquality || log.resolution || ""),
         videoSeconds: log.config?.videoSeconds || log.seconds || "",
+        videoGenerateAudio: log.config?.videoGenerateAudio || "true",
+        videoWatermark: log.config?.videoWatermark || "false",
     };
 }
 
-function buildLog({ prompt, model, config, references, videoReferences, durationMs, status, video, error }: { prompt: string; model: string; config: AiConfig; references: ReferenceImage[]; videoReferences: ReferenceVideo[]; durationMs: number; status: GenerationLog["status"]; video?: GeneratedVideo; error?: string }): GenerationLog {
+function buildLog({ prompt, model, config, references, videoReferences, audioReferences, durationMs, status, video, error }: { prompt: string; model: string; config: AiConfig; references: ReferenceImage[]; videoReferences: ReferenceVideo[]; audioReferences: ReferenceAudio[]; durationMs: number; status: GenerationLog["status"]; video?: GeneratedVideo; error?: string }): GenerationLog {
     const logConfig = {
         model: config.model,
         videoModel: config.videoModel,
         size: config.size,
         vquality: normalizeResolution(config.vquality),
         videoSeconds: config.videoSeconds,
+        videoGenerateAudio: config.videoGenerateAudio,
+        videoWatermark: config.videoWatermark,
     };
     return {
         id: nanoid(),
@@ -647,6 +730,7 @@ function buildLog({ prompt, model, config, references, videoReferences, duration
         config: logConfig,
         references,
         videoReferences,
+        audioReferences,
         durationMs,
         size: logConfig.size,
         resolution: logConfig.vquality,
@@ -658,17 +742,21 @@ function buildLog({ prompt, model, config, references, videoReferences, duration
 }
 
 function buildVideoConfig(config: AiConfig, model: string): AiConfig {
+    const seedance = isSeedanceVideoConfig({ ...config, model });
     return {
         ...config,
         model,
         videoModel: model,
-        size: normalizeVideoSize(config.size),
+        size: seedance ? normalizeSeedanceRatio(config.size) : normalizeVideoSize(config.size),
         videoSeconds: normalizeVideoSeconds(config.videoSeconds),
         vquality: normalizeResolution(config.vquality),
+        videoGenerateAudio: String(boolConfig(config.videoGenerateAudio, true)),
+        videoWatermark: String(boolConfig(config.videoWatermark, false)),
     };
 }
 
 function normalizeVideoSeconds(value: string) {
+    if (String(value).trim() === "-1") return "-1";
     const seconds = Math.floor(Number(value) || 6);
     return String(Math.max(1, Math.min(20, seconds)));
 }
diff --git a/web/src/components/video-settings-panel.tsx b/web/src/components/video-settings-panel.tsx
index cae6b3f..8e608b8 100644
--- a/web/src/components/video-settings-panel.tsx
+++ b/web/src/components/video-settings-panel.tsx
@@ -1,8 +1,10 @@
 "use client";
 
 import { type ReactNode } from "react";
+import { Switch } from "antd";
 
 import { ImageSettingsTheme } from "@/components/image-settings-panel";
+import { boolConfig, isSeedanceFastModel, isSeedanceVideoConfig, normalizeSeedanceDuration, normalizeSeedanceRatio, normalizeSeedanceResolution, seedanceDurationOptions, seedancePixelLabel, seedanceRatioOptions, seedanceResolutionOptions } from "@/lib/seedance-video";
 import { type CanvasTheme } from "@/lib/canvas-theme";
 import type { AiConfig } from "@/stores/use-config-store";
 
@@ -24,13 +26,17 @@ const secondOptions = [6, 10, 12, 16, 20];
 
 type VideoSettingsPanelProps = {
     config: AiConfig;
-    onConfigChange: (key: "vquality" | "size" | "videoSeconds", value: string) => void;
+    onConfigChange: (key: "vquality" | "size" | "videoSeconds" | "videoGenerateAudio" | "videoWatermark", value: string) => void;
     theme: CanvasTheme;
     showTitle?: boolean;
     className?: string;
 };
 
 export function VideoSettingsPanel({ config, onConfigChange, theme, showTitle = true, className = "w-[320px] space-y-4 rounded-2xl px-1 py-0.5" }: VideoSettingsPanelProps) {
+    if (isSeedanceVideoConfig(config)) {
+        return <SeedanceVideoSettingsPanel config={config} onConfigChange={onConfigChange} theme={theme} showTitle={showTitle} className={className} />;
+    }
+
     const seconds = config.videoSeconds || "6";
     const size = normalizeVideoSizeValue(config.size);
     const dimensions = readSizeDimensions(size);
@@ -96,16 +102,84 @@ export function VideoSettingsPanel({ config, onConfigChange, theme, showTitle =
     );
 }
 
+function SeedanceVideoSettingsPanel({ config, onConfigChange, theme, showTitle, className }: VideoSettingsPanelProps) {
+    const model = config.model || config.videoModel;
+    const resolution = normalizeSeedanceResolution(config.vquality, model);
+    const ratio = normalizeSeedanceRatio(config.size);
+    const duration = normalizeSeedanceDuration(config.videoSeconds);
+    const generateAudio = boolConfig(config.videoGenerateAudio, true);
+    const watermark = boolConfig(config.videoWatermark, false);
+
+    return (
+        <ImageSettingsTheme theme={theme}>
+            <div className={className} style={{ color: theme.node.text }} onMouseDown={(event) => event.stopPropagation()}>
+                {showTitle ? <div className="text-lg font-semibold">视频设置</div> : null}
+                <SettingGroup title="分辨率" color={theme.node.muted}>
+                    <div className="grid grid-cols-3 gap-2.5">
+                        {seedanceResolutionOptions.map((item) => {
+                            const disabled = item.value === "1080p" && isSeedanceFastModel(model);
+                            return (
+                                <OptionPill key={item.value} selected={resolution === item.value} disabled={disabled} theme={theme} onClick={() => onConfigChange("vquality", item.value)}>
+                                    {item.label}
+                                </OptionPill>
+                            );
+                        })}
+                    </div>
+                    {isSeedanceFastModel(model) ? <div className="text-[11px] leading-4 opacity-55">fast 模型不支持 1080p，会自动使用 720p。</div> : null}
+                </SettingGroup>
+                <SettingGroup title="比例" color={theme.node.muted}>
+                    <div className="grid grid-cols-3 gap-2.5">
+                        {seedanceRatioOptions.map((item) => (
+                            <button
+                                key={item.value}
+                                type="button"
+                                className="flex h-[68px] cursor-pointer flex-col items-center justify-center gap-1 rounded-xl border bg-transparent px-1 text-sm transition hover:opacity-80"
+                                style={{ borderColor: ratio === item.value ? theme.node.text : theme.node.stroke, color: theme.node.text }}
+                                onMouseDown={(event) => event.stopPropagation()}
+                                onClick={() => onConfigChange("size", item.value)}
+                            >
+                                <SizePreview width={ratioPreview(item.value).width} height={ratioPreview(item.value).height} color={theme.node.text} />
+                                <span>{item.label}</span>
+                                <span className="text-[10px] leading-none opacity-55">{item.value === "adaptive" ? "adaptive" : seedancePixelLabel(resolution, item.value)}</span>
+                            </button>
+                        ))}
+                    </div>
+                </SettingGroup>
+                <SettingGroup title="时长" color={theme.node.muted}>
+                    <div className="grid grid-cols-4 gap-2.5">
+                        {seedanceDurationOptions.map((value) => (
+                            <OptionPill key={value} selected={duration === value} theme={theme} onClick={() => onConfigChange("videoSeconds", String(value))}>
+                                {value === -1 ? "智能" : `${value}s`}
+                            </OptionPill>
+                        ))}
+                    </div>
+                    <NumberInput value={String(duration)} min={-1} max={15} theme={theme} onChange={(value) => onConfigChange("videoSeconds", value)} />
+                </SettingGroup>
+                <SettingGroup title="输出" color={theme.node.muted}>
+                    <div className="grid gap-2 rounded-xl border p-2.5" style={{ borderColor: theme.node.stroke }}>
+                        <SwitchRow label="生成声音" checked={generateAudio} theme={theme} onChange={(checked) => onConfigChange("videoGenerateAudio", String(checked))} />
+                        <SwitchRow label="添加水印" checked={watermark} theme={theme} onChange={(checked) => onConfigChange("videoWatermark", String(checked))} />
+                    </div>
+                </SettingGroup>
+            </div>
+        </ImageSettingsTheme>
+    );
+}
+
 export function videoResolutionLabel(value: string) {
     return `${normalizeVideoResolutionValue(value)}p`;
 }
 
 export function videoSizeLabel(value: string) {
+    const ratio = normalizeSeedanceRatio(value);
+    if (value === "adaptive" || value === "auto") return "自适应";
+    if (ratio === value) return seedanceRatioOptions.find((item) => item.value === ratio)?.label || ratio;
     const size = normalizeVideoSizeValue(value);
     return sizeOptions.find((item) => item.value === size)?.label || size;
 }
 
 export function videoSecondsLabel(value: string) {
+    if (String(value).trim() === "-1") return "智能";
     return `${value || "6"}s`;
 }
 
@@ -121,9 +195,9 @@ export function normalizeVideoResolutionValue(value: string) {
     return value.replace(/p$/i, "") || "720";
 }
 
-function OptionPill({ selected, theme, onClick, children }: { selected: boolean; theme: CanvasTheme; onClick: () => void; children: ReactNode }) {
+function OptionPill({ selected, disabled = false, theme, onClick, children }: { selected: boolean; disabled?: boolean; theme: CanvasTheme; onClick: () => void; children: ReactNode }) {
     return (
-        <button type="button" className="h-9 cursor-pointer rounded-full border px-2 text-sm transition hover:opacity-80" style={{ background: "transparent", borderColor: selected ? theme.node.text : theme.node.stroke, color: theme.node.text }} onMouseDown={(event) => event.stopPropagation()} onClick={onClick}>
+        <button type="button" disabled={disabled} className="h-9 cursor-pointer rounded-full border px-2 text-sm transition hover:opacity-80 disabled:cursor-not-allowed disabled:opacity-35" style={{ background: "transparent", borderColor: selected ? theme.node.text : theme.node.stroke, color: theme.node.text }} onMouseDown={(event) => event.stopPropagation()} onClick={onClick}>
             {children}
         </button>
     );
@@ -174,6 +248,29 @@ function SizePreview({ width, height, color }: { width: number; height: number;
     return <span className="rounded-[3px] border-2" style={{ width: previewWidth, height: previewHeight, borderColor: color }} />;
 }
 
+function ratioPreview(ratio: string) {
+    if (ratio === "9:16") return { width: 9, height: 16 };
+    if (ratio === "1:1") return { width: 1, height: 1 };
+    if (ratio === "4:3") return { width: 4, height: 3 };
+    if (ratio === "3:4") return { width: 3, height: 4 };
+    if (ratio === "21:9") return { width: 21, height: 9 };
+    if (ratio === "adaptive") return { width: 0, height: 0 };
+    return { width: 16, height: 9 };
+}
+
+function SwitchRow({ label, checked, theme, onChange }: { label: string; checked: boolean; theme: CanvasTheme; onChange: (checked: boolean) => void }) {
+    return (
+        <div className="flex h-8 items-center justify-between gap-3">
+            <span className="text-sm" style={{ color: theme.node.text }}>
+                {label}
+            </span>
+            <span onMouseDown={(event) => event.stopPropagation()}>
+                <Switch size="small" checked={checked} onChange={onChange} />
+            </span>
+        </div>
+    );
+}
+
 function readSizeDimensions(size: string) {
     if (size === "auto") return { width: 0, height: 0 };
     const match = size.match(/^(\d+)x(\d+)$/);
diff --git a/web/src/lib/seedance-video.ts b/web/src/lib/seedance-video.ts
new file mode 100644
index 0000000..35c00b9
--- /dev/null
+++ b/web/src/lib/seedance-video.ts
@@ -0,0 +1,125 @@
+import type { AiConfig } from "@/stores/use-config-store";
+
+export const SEEDANCE_REFERENCE_LIMITS = {
+    images: 9,
+    videos: 3,
+    audios: 3,
+    imageMaxBytes: 30 * 1024 * 1024,
+    videoMaxBytes: 50 * 1024 * 1024,
+    audioMaxBytes: 15 * 1024 * 1024,
+};
+
+export const seedanceResolutionOptions = [
+    { value: "480p", label: "480p" },
+    { value: "720p", label: "720p" },
+    { value: "1080p", label: "1080p" },
+] as const;
+
+export const seedanceRatioOptions = [
+    { value: "16:9", label: "横屏" },
+    { value: "9:16", label: "竖屏" },
+    { value: "1:1", label: "方形" },
+    { value: "4:3", label: "标准横屏" },
+    { value: "3:4", label: "标准竖屏" },
+    { value: "21:9", label: "宽银幕" },
+    { value: "adaptive", label: "自适应" },
+] as const;
+
+export const seedanceDurationOptions = [-1, 4, 5, 6, 8, 10, 12, 15] as const;
+
+const seedancePixels = {
+    "480p": {
+        "16:9": "864x496",
+        "4:3": "752x560",
+        "1:1": "640x640",
+        "3:4": "560x752",
+        "9:16": "496x864",
+        "21:9": "992x432",
+    },
+    "720p": {
+        "16:9": "1280x720",
+        "4:3": "1112x834",
+        "1:1": "960x960",
+        "3:4": "834x1112",
+        "9:16": "720x1280",
+        "21:9": "1470x630",
+    },
+    "1080p": {
+        "16:9": "1920x1080",
+        "4:3": "1664x1248",
+        "1:1": "1440x1440",
+        "3:4": "1248x1664",
+        "9:16": "1080x1920",
+        "21:9": "2206x946",
+    },
+} as const;
+
+export function isSeedanceVideoConfig(config: Pick<AiConfig, "model" | "videoModel" | "baseUrl">) {
+    return isSeedanceVideoModel(config.model || config.videoModel) || isArkPlanBaseUrl(config.baseUrl);
+}
+
+export function isSeedanceVideoModel(model: string) {
+    const value = model.toLowerCase();
+    return value.includes("seedance") || value.includes("doubao-seedance");
+}
+
+export function isSeedanceFastModel(model: string) {
+    const value = model.toLowerCase();
+    return isSeedanceVideoModel(value) && value.includes("fast");
+}
+
+export function isArkPlanBaseUrl(baseUrl: string) {
+    return baseUrl.toLowerCase().includes("ark.cn-beijing.volces.com/api/plan/v3") || baseUrl.toLowerCase().includes("/api/plan/v3");
+}
+
+export function normalizeSeedanceResolution(value: string, model = "") {
+    const normalized = normalizeResolutionToken(value);
+    if (isSeedanceFastModel(model) && normalized === "1080p") return "720p";
+    return seedanceResolutionOptions.some((item) => item.value === normalized) ? normalized : "720p";
+}
+
+export function normalizeResolutionToken(value: string) {
+    if (value === "low") return "480p";
+    if (value === "auto" || value === "high" || value === "medium") return "720p";
+    const resolution = String(value || "").replace(/p$/i, "") || "720";
+    return `${resolution}p`;
+}
+
+export function normalizeSeedanceDuration(value: string) {
+    if (String(value).trim() === "-1") return -1;
+    const seconds = Math.floor(Number(value) || 5);
+    return Math.max(4, Math.min(15, seconds));
+}
+
+export function normalizeSeedanceRatio(value: string) {
+    if (!value || value === "auto" || value === "adaptive") return "adaptive";
+    if (seedanceRatioOptions.some((item) => item.value === value)) return value;
+    const match = value.match(/^(\d+)x(\d+)$/);
+    if (!match) return "adaptive";
+    const width = Number(match[1]);
+    const height = Number(match[2]);
+    if (!width || !height) return "adaptive";
+    const ratio = width / height;
+    const options = [
+        ["16:9", 16 / 9],
+        ["4:3", 4 / 3],
+        ["1:1", 1],
+        ["3:4", 3 / 4],
+        ["9:16", 9 / 16],
+        ["21:9", 21 / 9],
+    ] as const;
+    return options.reduce((best, item) => (Math.abs(item[1] - ratio) < Math.abs(best[1] - ratio) ? item : best), options[0])[0];
+}
+
+export function seedancePixelLabel(resolution: string, ratio: string) {
+    const normalizedResolution = normalizeSeedanceResolution(resolution) as keyof typeof seedancePixels;
+    const normalizedRatio = normalizeSeedanceRatio(ratio) as keyof (typeof seedancePixels)[typeof normalizedResolution] | "adaptive";
+    if (normalizedRatio === "adaptive") return "自动匹配";
+    return seedancePixels[normalizedResolution][normalizedRatio] || "";
+}
+
+export function boolConfig(value: string | undefined, fallback: boolean) {
+    if (value === "true") return true;
+    if (value === "false") return false;
+    return fallback;
+}
diff --git a/web/src/services/api/video.ts b/web/src/services/api/video.ts
index 56ed12a..37522f8 100644
--- a/web/src/services/api/video.ts
+++ b/web/src/services/api/video.ts
@@ -3,10 +3,11 @@ import axios from "axios";
 import { dataUrlToFile } from "@/lib/image-utils";
 import { getMediaBlob, uploadMediaFile, type UploadedFile } from "@/services/file-storage";
 import { imageToDataUrl } from "@/services/image-storage";
+import { boolConfig, isSeedanceVideoConfig, normalizeSeedanceDuration, normalizeSeedanceRatio, normalizeSeedanceResolution, SEEDANCE_REFERENCE_LIMITS } from "@/lib/seedance-video";
 import { buildApiUrl, type AiConfig } from "@/stores/use-config-store";
 import { useUserStore } from "@/stores/use-user-store";
 import type { ReferenceImage } from "@/types/image";
-import type { ReferenceVideo } from "@/types/media";
+import type { ReferenceAudio, ReferenceVideo } from "@/types/media";
 
 type VideoResponse = { id: string; status?: string; error?: { message?: string } };
 type ApiVideoResponse = VideoResponse | { code?: number; data?: VideoResponse | null; msg?: string };
@@ -42,14 +43,14 @@ function refreshRemoteUser(config: AiConfig) {
     if (config.channelMode === "remote") void useUserStore.getState().hydrateUser();
 }
 
-export async function requestVideoGeneration(config: AiConfig, prompt: string, references: ReferenceImage[] = [], videoReferences: ReferenceVideo[] = []): Promise<VideoGenerationResult> {
+export async function requestVideoGeneration(config: AiConfig, prompt: string, references: ReferenceImage[] = [], videoReferences: ReferenceVideo[] = [], audioReferences: ReferenceAudio[] = []): Promise<VideoGenerationResult> {
     const model = (config.model || config.videoModel).trim();
     assertVideoConfig(config, model);
-    if (isSeedanceConfig(config, model)) {
-        return requestSeedanceGeneration(config, model, prompt, references, videoReferences);
+    if (isSeedanceVideoConfig({ ...config, model })) {
+        return requestSeedanceGeneration(config, model, prompt, references, videoReferences, audioReferences);
     }
-    if (videoReferences.length) {
-        throw new Error("当前视频接口不支持参考视频，请切换到 Seedance 2.0 / 火山 Agent Plan 模型，或移除参考视频");
+    if (videoReferences.length || audioReferences.length) {
+        throw new Error("当前视频接口不支持参考视频或参考音频，请切换到 Seedance 2.0 / 火山 Agent Plan 模型，或移除参考素材");
     }
     return requestOpenAIVideoGeneration(config, model, prompt, references);
 }
@@ -89,16 +90,22 @@ async function requestOpenAIVideoGeneration(config: AiConfig, model: string, pro
     }
 }
 
-async function requestSeedanceGeneration(config: AiConfig, model: string, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[]) {
-    const content = await buildSeedanceContent(config, prompt, references, videoReferences);
-    if (!content.length) throw new Error("请输入视频提示词，或连接参考图片/视频");
+async function requestSeedanceGeneration(config: AiConfig, model: string, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[], audioReferences: ReferenceAudio[]) {
+    if (audioReferences.length && !references.length && !videoReferences.length) {
+        throw new Error("Seedance 参考音频不能单独使用，请同时添加参考图或参考视频");
+    }
+    assertSeedanceVideoReferences(videoReferences);
+    assertSeedanceAudioReferences(audioReferences);
+    const content = await buildSeedanceContent(config, prompt, references, videoReferences, audioReferences);
+    if (!content.length) throw new Error("请输入视频提示词，或连接参考图片/视频/音频");
     const payload = {
         model,
         content,
         ratio: normalizeSeedanceRatio(config.size),
-        resolution: normalizeVideoResolution(config.vquality),
+        resolution: normalizeSeedanceResolution(config.vquality, model),
         duration: normalizeSeedanceDuration(config.videoSeconds),
-        watermark: false,
+        generate_audio: boolConfig(config.videoGenerateAudio, true),
+        watermark: boolConfig(config.videoWatermark, false),
     };
 
     try {
@@ -122,21 +129,44 @@ async function requestSeedanceGeneration(config: AiConfig, model: string, prompt
     }
 }
 
+function assertSeedanceVideoReferences(videoReferences: ReferenceVideo[]) {
+    let total = 0;
+    for (const video of videoReferences) {
+        if (!video.durationMs) continue;
+        if (video.durationMs < 2000 || video.durationMs > 15000) throw new Error("Seedance 参考视频单个时长需要在 2-15 秒之间");
+        total += video.durationMs;
+    }
+    if (total > 15000) throw new Error("Seedance 参考视频总时长不能超过 15 秒");
+}
+
+function assertSeedanceAudioReferences(audioReferences: ReferenceAudio[]) {
+    let total = 0;
+    for (const audio of audioReferences) {
+        if (!audio.durationMs) continue;
+        if (audio.durationMs < 2000 || audio.durationMs > 15000) throw new Error("Seedance 参考音频单个时长需要在 2-15 秒之间");
+        total += audio.durationMs;
+    }
+    if (total > 15000) throw new Error("Seedance 参考音频总时长不能超过 15 秒");
+}
+
 function seedanceApiUrl(config: AiConfig, taskId?: string) {
     if (config.channelMode === "remote") return taskId ? `/api/v1/videos/${encodeURIComponent(taskId)}` : "/api/v1/videos";
     return buildApiUrl(config.baseUrl, `/contents/generations/tasks${taskId ? `/${encodeURIComponent(taskId)}` : ""}`);
 }
 
-async function buildSeedanceContent(config: AiConfig, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[]) {
+async function buildSeedanceContent(config: AiConfig, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[], audioReferences: ReferenceAudio[]) {
     const content: Array<Record<string, unknown>> = [];
     const text = prompt.trim();
     if (text) content.push({ type: "text", text });
-    for (const image of references.slice(0, 7)) {
+    for (const image of references.slice(0, SEEDANCE_REFERENCE_LIMITS.images)) {
         content.push({ type: "image_url", image_url: { url: await resolveSeedanceImageUrl(config, image) }, role: "reference_image" });
     }
-    for (const video of videoReferences.slice(0, 3)) {
+    for (const video of videoReferences.slice(0, SEEDANCE_REFERENCE_LIMITS.videos)) {
         content.push({ type: "video_url", video_url: { url: await resolveSeedanceVideoUrl(video) }, role: "reference_video" });
     }
+    for (const audio of audioReferences.slice(0, SEEDANCE_REFERENCE_LIMITS.audios)) {
+        content.push({ type: "audio_url", audio_url: { url: await resolveSeedanceAudioUrl(audio) }, role: "reference_audio" });
+    }
     return content;
 }
 
@@ -161,6 +191,16 @@ async function resolveSeedanceVideoUrl(video: ReferenceVideo) {
     return uploadReferenceMedia(file);
 }
 
+async function resolveSeedanceAudioUrl(audio: ReferenceAudio) {
+    if (isPublicMediaUrl(audio.url) || audio.url.startsWith("asset://")) return audio.url;
+    let blob: Blob | null = null;
+    if (audio.storageKey) blob = await getMediaBlob(audio.storageKey);
+    if (!blob && audio.url?.startsWith("blob:")) blob = await (await fetch(audio.url)).blob();
+    if (!blob) throw new Error("参考音频必须是公网 URL、素材 ID，或本地已保存的音频");
+    const file = new File([blob], audio.name || "reference-audio.mp3", { type: audio.type || blob.type || "audio/mpeg" });
+    return uploadReferenceMedia(file);
+}
+
 async function uploadReferenceMedia(file: File) {
     const token = useUserStore.getState().token;
     if (!token) throw new Error("使用本地参考素材需要先登录，并在服务端配置 PUBLIC_BASE_URL");
@@ -182,11 +222,6 @@ async function videoResultFromUrl(url: string): Promise<VideoGenerationResult> {
     }
 }
 
-function isSeedanceConfig(config: AiConfig, model: string) {
-    const value = `${model} ${config.baseUrl}`.toLowerCase();
-    return value.includes("seedance") || value.includes("doubao-seedance") || value.includes("ark.cn-beijing.volces.com/api/plan/v3");
-}
-
 function assertVideoConfig(config: AiConfig, model: string) {
     if (!model) throw new Error("请先配置视频模型");
     if (config.channelMode === "local" && !config.baseUrl.trim()) throw new Error("请先配置 Base URL");
@@ -198,11 +233,6 @@ function normalizeVideoSeconds(value: string) {
     return String(Math.max(1, Math.min(20, seconds)));
 }
 
-function normalizeSeedanceDuration(value: string) {
-    const seconds = Math.floor(Number(value) || 5);
-    return Math.max(4, Math.min(15, seconds));
-}
-
 function normalizeVideoSize(value: string) {
     if (value === "auto") return null;
     const size = value || "1280x720";
@@ -210,26 +240,6 @@ function normalizeVideoSize(value: string) {
     return ["9:16", "2:3", "3:4"].includes(size) ? "720x1280" : "1280x720";
 }
 
-function normalizeSeedanceRatio(value: string) {
-    if (!value || value === "auto") return "adaptive";
-    if (["16:9", "4:3", "1:1", "3:4", "9:16", "21:9", "adaptive"].includes(value)) return value;
-    const match = value.match(/^(\d+)x(\d+)$/);
-    if (!match) return "adaptive";
-    const width = Number(match[1]);
-    const height = Number(match[2]);
-    if (!width || !height) return "adaptive";
-    const ratio = width / height;
-    const options = [
-        ["16:9", 16 / 9],
-        ["4:3", 4 / 3],
-        ["1:1", 1],
-        ["3:4", 3 / 4],
-        ["9:16", 9 / 16],
-        ["21:9", 21 / 9],
-    ] as const;
-    return options.reduce((best, item) => (Math.abs(item[1] - ratio) < Math.abs(best[1] - ratio) ? item : best), options[0])[0];
-}
-
 function normalizeVideoResolution(value: string) {
     if (value === "low") return "480p";
     if (value === "auto" || value === "high" || value === "medium") return "720p";
diff --git a/web/src/services/file-storage.ts b/web/src/services/file-storage.ts
index 6a1eed6..60ee07a 100644
--- a/web/src/services/file-storage.ts
+++ b/web/src/services/file-storage.ts
@@ -3,7 +3,7 @@
 import localforage from "localforage";
 import { nanoid } from "nanoid";
 
-export type UploadedFile = { url: string; storageKey: string; bytes: number; mimeType: string; width?: number; height?: number };
+export type UploadedFile = { url: string; storageKey: string; bytes: number; mimeType: string; width?: number; height?: number; durationMs?: number };
 
 const store = localforage.createInstance({ name: "infinite-canvas", storeName: "media_files" });
 const objectUrls = new Map<string, string>();
@@ -14,7 +14,7 @@ export async function uploadMediaFile(input: string | Blob, prefix = "file"): Pr
     await store.setItem(storageKey, blob);
     const url = URL.createObjectURL(blob);
     objectUrls.set(storageKey, url);
-    const meta = blob.type.startsWith("video/") ? await readVideoMeta(url) : {};
+    const meta = blob.type.startsWith("video/") ? await readVideoMeta(url) : blob.type.startsWith("audio/") ? await readAudioMeta(url) : {};
     return { url, storageKey, bytes: blob.size, mimeType: blob.type || "application/octet-stream", ...meta };
 }
 
@@ -68,11 +68,21 @@ export function collectMediaStorageKeys(value: unknown, keys = new Set<string>()
 }
 
 function readVideoMeta(url: string) {
-    return new Promise<{ width: number; height: number }>((resolve) => {
+    return new Promise<{ width: number; height: number; durationMs?: number }>((resolve) => {
         const video = document.createElement("video");
-        const done = () => resolve({ width: video.videoWidth || 1280, height: video.videoHeight || 720 });
+        const done = () => resolve({ width: video.videoWidth || 1280, height: video.videoHeight || 720, durationMs: Number.isFinite(video.duration) ? Math.round(video.duration * 1000) : undefined });
         video.onloadedmetadata = done;
         video.onerror = done;
         video.src = url;
     });
 }
+
+function readAudioMeta(url: string) {
+    return new Promise<{ durationMs?: number }>((resolve) => {
+        const audio = document.createElement("audio");
+        const done = () => resolve({ durationMs: Number.isFinite(audio.duration) ? Math.round(audio.duration * 1000) : undefined });
+        audio.onloadedmetadata = done;
+        audio.onerror = done;
+        audio.src = url;
+    });
+}
diff --git a/web/src/stores/use-config-store.ts b/web/src/stores/use-config-store.ts
index 49e0af4..625fecb 100644
--- a/web/src/stores/use-config-store.ts
+++ b/web/src/stores/use-config-store.ts
@@ -17,6 +17,8 @@ export type AiConfig = {
     textModel: string;
     videoSeconds: string;
     vquality: string;
+    videoGenerateAudio: string;
+    videoWatermark: string;
     systemPrompt: string;
     models: string[];
     quality: string;
@@ -36,6 +38,8 @@ export const defaultConfig: AiConfig = {
     textModel: "gpt-5.5",
     videoSeconds: "6",
     vquality: "720",
+    videoGenerateAudio: "true",
+    videoWatermark: "false",
     systemPrompt: "",
     models: [],
     quality: "auto",
@@ -61,19 +65,44 @@ function resolveEffectiveConfig(config: AiConfig, modelChannel: AdminPublicSetti
     const channelMode = modelChannel?.allowCustomChannel ? config.channelMode : "remote";
     if (channelMode === "local" || !modelChannel) return { ...config, channelMode };
     const models = modelChannel.availableModels;
-    const fallbackModel = modelChannel.defaultModel || models[0] || "";
+    const fallbackTextModel = validDefault(modelChannel.defaultTextModel, models) || preferredModel(models, isTextModelName);
+    const fallbackModel = validDefault(modelChannel.defaultModel, models) || fallbackTextModel || models[0] || "";
+    const fallbackImageModel = validDefault(modelChannel.defaultImageModel, models) || preferredModel(models, isImageModelName) || fallbackModel;
+    const fallbackVideoModel = validDefault(modelChannel.defaultVideoModel, models) || preferredModel(models, isVideoModelName) || fallbackModel;
     return {
         ...config,
         channelMode,
         models,
         model: models.includes(config.model) ? config.model : fallbackModel,
-        imageModel: models.includes(config.imageModel) ? config.imageModel : modelChannel.defaultImageModel || fallbackModel,
-        videoModel: models.includes(config.videoModel) ? config.videoModel : modelChannel.defaultVideoModel || fallbackModel,
-        textModel: models.includes(config.textModel) ? config.textModel : modelChannel.defaultTextModel || fallbackModel,
+        imageModel: models.includes(config.imageModel) ? config.imageModel : fallbackImageModel,
+        videoModel: models.includes(config.videoModel) ? config.videoModel : fallbackVideoModel,
+        textModel: models.includes(config.textModel) ? config.textModel : fallbackTextModel || fallbackModel,
         systemPrompt: modelChannel.systemPrompt,
     };
 }
 
+function validDefault(model: string, models: string[]) {
+    return models.includes(model) ? model : "";
+}
+
+function preferredModel(models: string[], predicate: (model: string) => boolean) {
+    return models.find(predicate) || "";
+}
+
+function isVideoModelName(model: string) {
+    const value = model.toLowerCase();
+    return value.includes("seedance") || value.includes("video");
+}
+
+function isImageModelName(model: string) {
+    const value = model.toLowerCase();
+    return value.includes("seedream") || value.includes("gpt-image") || value.includes("image");
+}
+
+function isTextModelName(model: string) {
+    return !isImageModelName(model) && !isVideoModelName(model);
+}
+
 function isAiConfigReady(config: AiConfig, model: string) {
     return Boolean(model.trim()) && (config.channelMode === "remote" || Boolean(config.baseUrl.trim() && config.apiKey.trim()));
 }
@@ -112,7 +141,7 @@ export const useConfigStore = create<ConfigStore>()(
             partialize: (state) => ({ config: state.config }),
             merge: (persisted, current) => {
                 const config = { ...defaultConfig, ...((persisted as Partial<ConfigStore>).config || {}) };
-                return { ...current, config: { ...config, channelMode: config.channelMode || "remote", imageModel: config.imageModel || config.model, videoModel: config.videoModel || "grok-imagine-video", textModel: config.textModel || config.model, videoSeconds: config.videoSeconds || "6", vquality: config.vquality || "720" } };
+                return { ...current, config: { ...config, channelMode: config.channelMode || "remote", imageModel: config.imageModel || config.model, videoModel: config.videoModel || "grok-imagine-video", textModel: config.textModel || config.model, videoSeconds: config.videoSeconds || "6", vquality: config.vquality || "720", videoGenerateAudio: config.videoGenerateAudio || "true", videoWatermark: config.videoWatermark || "false" } };
             },
         },
     ),
diff --git a/web/src/types/media.ts b/web/src/types/media.ts
index 6d95955..84d4e2d 100644
--- a/web/src/types/media.ts
+++ b/web/src/types/media.ts
@@ -4,4 +4,14 @@ export type ReferenceVideo = {
     type: string;
     url: string;
     storageKey?: string;
+    durationMs?: number;
+};
+
+export type ReferenceAudio = {
+    id: string;
+    name: string;
+    type: string;
+    url: string;
+    storageKey?: string;
+    durationMs?: number;
 };

From 74ea4b81f2547f0b7a9f81f8eea278ce5e07f376 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 21:55:05 +0800
Subject: [PATCH 10/14] fix: improve seedance reference handling

---
 docs/pending-test.md                          |  1 +
 handler/ai.go                                 | 52 ++++++++++++++++++-
 handler/ai_test.go                            | 20 +++++++
 .../components/canvas-config-node-panel.tsx   |  7 +--
 .../components/canvas-node-generation.ts      |  4 ++
 web/src/app/(user)/video/page.tsx             | 43 ++++++++++++---
 web/src/lib/seedance-video.ts                 | 43 +++++++++++++++
 web/src/services/api/video.ts                 |  6 ++-
 web/src/types/media.ts                        |  3 ++
 9 files changed, 165 insertions(+), 14 deletions(-)
 create mode 100644 handler/ai_test.go

diff --git a/docs/pending-test.md b/docs/pending-test.md
index e313d2b..24a039d 100644
--- a/docs/pending-test.md
+++ b/docs/pending-test.md
@@ -1,5 +1,6 @@
 # 待测试
 
+- Seedance 参考素材失败原因排查：后端会把火山上游错误摘要返回给前端；`/video` 和画布视频生成会按 `图片1/视频1/音频1` 自动编号参考素材，并在实际请求提示词中注入编号说明；参考视频会在请求前校验大小、时长、宽高、宽高比和像素总量。
 - 画布项目导出改为下载 `.zip` 压缩包，包内包含 `projects.json` 和当前画布引用到的本地图片、视频文件，避免只导出 JSON 时丢失媒体内容。
 - 画布库支持多选后一键导出多个画布项目，导出的压缩包可一次恢复多个项目。
 - 画布项目导入改为读取新版 `.zip` 压缩包，会先按 `projects.json` 中的文件映射恢复图片、视频到本地存储，再插入画布项目，导入成功后仍停留在画布库。
diff --git a/handler/ai.go b/handler/ai.go
index 045936c..c569c0a 100644
--- a/handler/ai.go
+++ b/handler/ai.go
@@ -119,12 +119,12 @@ func copyAIResponse(w http.ResponseWriter, request *http.Request, onFailure func
 	defer response.Body.Close()
 
 	if response.StatusCode >= http.StatusBadRequest {
-		_, _ = io.Copy(io.Discard, io.LimitReader(response.Body, 4096))
+		body, _ := io.ReadAll(io.LimitReader(response.Body, 4096))
 		log.Printf("AI upstream error: url=%s status=%d", request.URL.String(), response.StatusCode)
 		if onFailure != nil {
 			onFailure()
 		}
-		Fail(w, aiStatusMessage(response.StatusCode))
+		Fail(w, aiUpstreamStatusMessage(response.StatusCode, body))
 		return
 	}
 
@@ -239,6 +239,54 @@ func aiStatusMessage(statusCode int) string {
 	}
 }
 
+func aiUpstreamStatusMessage(statusCode int, body []byte) string {
+	base := aiStatusMessage(statusCode)
+	detail := aiUpstreamErrorDetail(body)
+	if detail == "" {
+		return base
+	}
+	return base + "：" + detail
+}
+
+func aiUpstreamErrorDetail(body []byte) string {
+	text := strings.TrimSpace(string(body))
+	if text == "" {
+		return ""
+	}
+	var payload struct {
+		Msg     string `json:"msg"`
+		Message string `json:"message"`
+		Error   struct {
+			Code    string `json:"code"`
+			Message string `json:"message"`
+		} `json:"error"`
+	}
+	if err := json.Unmarshal(body, &payload); err == nil {
+		if payload.Error.Message != "" {
+			if payload.Error.Code != "" {
+				return safeUpstreamText(payload.Error.Code + " " + payload.Error.Message)
+			}
+			return safeUpstreamText(payload.Error.Message)
+		}
+		if payload.Msg != "" {
+			return safeUpstreamText(payload.Msg)
+		}
+		if payload.Message != "" {
+			return safeUpstreamText(payload.Message)
+		}
+	}
+	return safeUpstreamText(text)
+}
+
+func safeUpstreamText(text string) string {
+	text = strings.Join(strings.Fields(strings.TrimSpace(text)), " ")
+	runes := []rune(text)
+	if len(runes) > 300 {
+		return string(runes[:300]) + "..."
+	}
+	return text
+}
+
 type aiError struct {
 	message string
 }
diff --git a/handler/ai_test.go b/handler/ai_test.go
new file mode 100644
index 0000000..2c4d8d4
--- /dev/null
+++ b/handler/ai_test.go
@@ -0,0 +1,20 @@
+package handler
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestAIUpstreamErrorDetail(t *testing.T) {
+	got := aiUpstreamErrorDetail([]byte(`{"error":{"code":"InvalidParameter","message":"reference video fps is invalid"}}`))
+	if got != "InvalidParameter reference video fps is invalid" {
+		t.Fatalf("detail = %q", got)
+	}
+}
+
+func TestSafeUpstreamTextTruncates(t *testing.T) {
+	got := safeUpstreamText(strings.Repeat("错", 320))
+	if len([]rune(got)) != 303 {
+		t.Fatalf("truncated rune length = %d", len([]rune(got)))
+	}
+}
diff --git a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
index 8966229..f910ed7 100644
--- a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
@@ -9,6 +9,7 @@ import { ModelPicker } from "@/components/model-picker";
 import { defaultConfig, useConfigStore, useEffectiveConfig, type AiConfig } from "@/stores/use-config-store";
 import { CreditSymbol, requestCreditCost } from "@/constant/credits";
 import { canvasThemes } from "@/lib/canvas-theme";
+import { seedanceReferenceLabel } from "@/lib/seedance-video";
 import { useThemeStore } from "@/stores/use-theme-store";
 import { CanvasImageSettingsPopover } from "./canvas-image-settings-popover";
 import { CanvasVideoSettingsPopover } from "./canvas-video-settings-popover";
@@ -303,7 +304,7 @@ function ImageSortCard({
         <div className="w-24 shrink-0 overflow-hidden rounded-lg border" style={{ background: theme.node.fill, borderColor: theme.node.stroke }}>
             <div className="relative">
                 <img src={input.image.dataUrl} alt={input.title} className="aspect-square w-full object-cover" />
-                <span className="absolute left-1 top-1 rounded bg-black/50 px-1 py-0.5 text-[9px] font-medium text-white">{imageIndex + 1}</span>
+                <span className="absolute left-1 top-1 rounded bg-black/50 px-1 py-0.5 text-[9px] font-medium text-white">{seedanceReferenceLabel("image", imageIndex)}</span>
                 <HorizontalOrderButtons index={imageIndex} total={imageTotal} onMove={(offset) => onMove(input, offset)} />
             </div>
         </div>
@@ -328,7 +329,7 @@ function VideoSortCard({
         <div className="w-32 shrink-0 overflow-hidden rounded-lg border" style={{ background: theme.node.fill, borderColor: theme.node.stroke }}>
             <div className="relative">
                 <video src={input.video.url} className="aspect-video w-full bg-black object-cover" muted preload="metadata" />
-                <span className="absolute left-1 top-1 rounded bg-black/50 px-1 py-0.5 text-[9px] font-medium text-white">{videoIndex + 1}</span>
+                <span className="absolute left-1 top-1 rounded bg-black/50 px-1 py-0.5 text-[9px] font-medium text-white">{seedanceReferenceLabel("video", videoIndex)}</span>
                 <HorizontalOrderButtons index={videoIndex} total={videoTotal} onMove={(offset) => onMove(input, offset)} />
             </div>
         </div>
@@ -358,7 +359,7 @@ function AudioSortCard({
             <audio src={input.audio.url} controls className="h-8 w-full" preload="metadata" />
             <div className="mt-1 flex justify-between">
                 <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !p-0" icon={<ArrowLeft className="size-3" />} disabled={audioIndex <= 0} onClick={() => onMove(input, -1)} />
-                <span className="text-[10px] opacity-45">{audioIndex + 1}</span>
+                <span className="text-[10px] opacity-45">{seedanceReferenceLabel("audio", audioIndex)}</span>
                 <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !p-0" icon={<ArrowRight className="size-3" />} disabled={audioIndex >= audioTotal - 1} onClick={() => onMove(input, 1)} />
             </div>
         </div>
diff --git a/web/src/app/(user)/canvas/components/canvas-node-generation.ts b/web/src/app/(user)/canvas/components/canvas-node-generation.ts
index 3a3f1a4..abde93a 100644
--- a/web/src/app/(user)/canvas/components/canvas-node-generation.ts
+++ b/web/src/app/(user)/canvas/components/canvas-node-generation.ts
@@ -102,6 +102,10 @@ function readReferenceVideo(node: CanvasNodeData): ReferenceVideo | null {
         type: node.metadata.mimeType || "video/mp4",
         url: node.metadata.content,
         storageKey: node.metadata.storageKey,
+        bytes: node.metadata.bytes,
+        width: node.metadata.naturalWidth,
+        height: node.metadata.naturalHeight,
+        durationMs: node.metadata.durationMs,
     };
 }
 
diff --git a/web/src/app/(user)/video/page.tsx b/web/src/app/(user)/video/page.tsx
index 7a6b009..2cbd3c7 100644
--- a/web/src/app/(user)/video/page.tsx
+++ b/web/src/app/(user)/video/page.tsx
@@ -1,6 +1,6 @@
 "use client";
 
-import { BookOpen, CheckSquare, ClipboardPaste, Download, FolderPlus, History, LoaderCircle, Music2, Plus, SlidersHorizontal, Sparkles, Trash2, Upload, VideoIcon } from "lucide-react";
+import { ArrowLeft, ArrowRight, BookOpen, CheckSquare, ClipboardPaste, Download, FolderPlus, History, LoaderCircle, Music2, Plus, SlidersHorizontal, Sparkles, Trash2, Upload, VideoIcon } from "lucide-react";
 import { useEffect, useRef, useState } from "react";
 import { App, Button, Checkbox, Drawer, Empty, Input, Modal, Tag, Typography } from "antd";
 import localforage from "localforage";
@@ -13,7 +13,7 @@ import { PromptSelectDialog } from "@/components/prompts/prompt-select-dialog";
 import { VideoSettingsPanel, normalizeVideoResolutionValue, normalizeVideoSizeValue, videoSizeLabel } from "@/components/video-settings-panel";
 import { canvasThemes } from "@/lib/canvas-theme";
 import { formatBytes, formatDuration } from "@/lib/image-utils";
-import { boolConfig, isSeedanceVideoConfig, normalizeSeedanceRatio, SEEDANCE_REFERENCE_LIMITS } from "@/lib/seedance-video";
+import { boolConfig, isSeedanceVideoConfig, normalizeSeedanceRatio, seedanceReferenceLabel, seedanceVideoReferenceError, seedanceVideoReferenceHint, SEEDANCE_REFERENCE_LIMITS } from "@/lib/seedance-video";
 import { deleteStoredMedia, resolveMediaUrl, uploadMediaFile } from "@/services/file-storage";
 import { resolveImageUrl, uploadImage } from "@/services/image-storage";
 import { requestVideoGeneration, storeGeneratedVideo } from "@/services/api/video";
@@ -126,7 +126,7 @@ export default function VideoPage() {
         const nextVideoReferences = await Promise.all(
             videoFiles.map(async (file) => {
                 const video = await uploadMediaFile(file, "video-reference");
-                return { id: nanoid(), name: file.name, type: video.mimeType, url: video.url, storageKey: video.storageKey, durationMs: video.durationMs };
+                return { id: nanoid(), name: file.name, type: video.mimeType, url: video.url, storageKey: video.storageKey, bytes: video.bytes, width: video.width, height: video.height, durationMs: video.durationMs };
             }),
         );
         const nextAudioReferences = filterAudioReferencesByDuration(
@@ -209,6 +209,11 @@ export default function VideoPage() {
             openConfigDialog(true);
             return null;
         }
+        const videoReferenceError = seedanceVideoReferenceError(videoReferences);
+        if (videoReferenceError) {
+            message.error(`${videoReferenceError}。${seedanceVideoReferenceHint}`);
+            return null;
+        }
         return { text, config: buildVideoConfig(effectiveConfig, model), references: [...references], videoReferences: [...videoReferences], audioReferences: [...audioReferences] };
     };
 
@@ -240,7 +245,7 @@ export default function VideoPage() {
             const stored = await uploadImage(payload.dataUrl);
             setReferences((value) => [...value, { id: nanoid(), name: payload.title, type: stored.mimeType, dataUrl: stored.url, storageKey: stored.storageKey }].slice(0, SEEDANCE_REFERENCE_LIMITS.images));
         } else if (payload.kind === "video") {
-            setVideoReferences((value) => [...value, { id: nanoid(), name: payload.title, type: "video/mp4", url: payload.url, storageKey: payload.storageKey }].slice(0, SEEDANCE_REFERENCE_LIMITS.videos));
+            setVideoReferences((value) => [...value, { id: nanoid(), name: payload.title, type: "video/mp4", url: payload.url, storageKey: payload.storageKey, width: payload.width, height: payload.height }].slice(0, SEEDANCE_REFERENCE_LIMITS.videos));
         }
         setAssetPickerOpen(false);
     };
@@ -343,9 +348,11 @@ export default function VideoPage() {
                                     </div>
                                 </div>
                                 <div className="hover-scrollbar hover-scrollbar-hint flex min-h-24 w-full min-w-0 max-w-full gap-2 overflow-x-scroll overflow-y-hidden rounded-lg border border-dashed border-stone-300 p-2 pb-3 overscroll-x-contain dark:border-stone-700">
-                                    {references.map((item) => (
+                                    {references.map((item, index) => (
                                         <div key={item.id} className="group relative size-20 shrink-0 overflow-hidden rounded-md border border-stone-200 dark:border-stone-800">
                                             <img src={item.dataUrl} alt={item.name} className="size-full object-cover" />
+                                            <span className="absolute left-1 top-1 rounded bg-black/60 px-1.5 py-0.5 text-[10px] font-medium text-white">{seedanceReferenceLabel("image", index)}</span>
+                                            <ReferenceOrderButtons index={index} total={references.length} onMove={(offset) => setReferences((value) => moveListItem(value, index, offset))} />
                                             <button type="button" className="absolute right-1 top-1 hidden size-6 items-center justify-center rounded bg-black/60 text-white group-hover:flex" onClick={() => setReferences((value) => value.filter((ref) => ref.id !== item.id))} aria-label="移除参考图">
                                                 <Trash2 className="size-3.5" />
                                             </button>
@@ -363,9 +370,11 @@ export default function VideoPage() {
                                     </Button>
                                 </div>
                                 <div className="hover-scrollbar hover-scrollbar-hint flex min-h-24 w-full min-w-0 max-w-full gap-2 overflow-x-scroll overflow-y-hidden rounded-lg border border-dashed border-stone-300 p-2 pb-3 overscroll-x-contain dark:border-stone-700">
-                                    {videoReferences.map((item) => (
+                                    {videoReferences.map((item, index) => (
                                         <div key={item.id} className="group relative h-20 w-32 shrink-0 overflow-hidden rounded-md border border-stone-200 bg-black dark:border-stone-800">
                                             <video src={item.url} className="size-full object-cover" muted preload="metadata" />
+                                            <span className="absolute left-1 top-1 rounded bg-black/60 px-1.5 py-0.5 text-[10px] font-medium text-white">{seedanceReferenceLabel("video", index)}</span>
+                                            <ReferenceOrderButtons index={index} total={videoReferences.length} onMove={(offset) => setVideoReferences((value) => moveListItem(value, index, offset))} />
                                             <button type="button" className="absolute right-1 top-1 hidden size-6 items-center justify-center rounded bg-black/60 text-white group-hover:flex" onClick={() => setVideoReferences((value) => value.filter((ref) => ref.id !== item.id))} aria-label="移除参考视频">
                                                 <Trash2 className="size-3.5" />
                                             </button>
@@ -383,13 +392,15 @@ export default function VideoPage() {
                                     </Button>
                                 </div>
                                 <div className="hover-scrollbar hover-scrollbar-hint flex min-h-24 w-full min-w-0 max-w-full gap-2 overflow-x-scroll overflow-y-hidden rounded-lg border border-dashed border-stone-300 p-2 pb-3 overscroll-x-contain dark:border-stone-700">
-                                    {audioReferences.map((item) => (
+                                    {audioReferences.map((item, index) => (
                                         <div key={item.id} className="group relative flex h-20 w-48 shrink-0 flex-col justify-center gap-2 rounded-md border border-stone-200 bg-stone-50 px-2 dark:border-stone-800 dark:bg-stone-900">
                                             <div className="flex min-w-0 items-center gap-2 text-xs text-stone-500 dark:text-stone-400">
                                                 <Music2 className="size-4 shrink-0" />
+                                                <span className="shrink-0 rounded bg-stone-200 px-1 text-[10px] text-stone-700 dark:bg-stone-800 dark:text-stone-200">{seedanceReferenceLabel("audio", index)}</span>
                                                 <span className="truncate">{item.name}</span>
                                             </div>
                                             <audio src={item.url} controls className="h-8 w-full" preload="metadata" />
+                                            <ReferenceOrderButtons index={index} total={audioReferences.length} onMove={(offset) => setAudioReferences((value) => moveListItem(value, index, offset))} />
                                             <button type="button" className="absolute right-1 top-1 hidden size-6 items-center justify-center rounded bg-black/60 text-white group-hover:flex" onClick={() => setAudioReferences((value) => value.filter((ref) => ref.id !== item.id))} aria-label="移除参考音频">
                                                 <Trash2 className="size-3.5" />
                                             </button>
@@ -698,6 +709,24 @@ function filterAudioReferencesByDuration(existing: ReferenceAudio[], next: Refer
     return accepted;
 }
 
+function moveListItem<T>(items: T[], index: number, offset: number) {
+    const targetIndex = index + offset;
+    if (targetIndex < 0 || targetIndex >= items.length) return items;
+    const next = [...items];
+    [next[index], next[targetIndex]] = [next[targetIndex], next[index]];
+    return next;
+}
+
+function ReferenceOrderButtons({ index, total, onMove }: { index: number; total: number; onMove: (offset: number) => void }) {
+    if (total <= 1) return null;
+    return (
+        <div className="absolute inset-x-1 bottom-1 flex justify-between">
+            <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !bg-white/85 !p-0 !shadow-sm" icon={<ArrowLeft className="size-3" />} disabled={index <= 0} onClick={() => onMove(-1)} />
+            <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !bg-white/85 !p-0 !shadow-sm" icon={<ArrowRight className="size-3" />} disabled={index >= total - 1} onClick={() => onMove(1)} />
+        </div>
+    );
+}
+
 function normalizeLogConfig(log: Partial<GenerationLog>): GenerationLogConfig {
     return {
         model: log.config?.model || log.model || "",
diff --git a/web/src/lib/seedance-video.ts b/web/src/lib/seedance-video.ts
index 35c00b9..4194396 100644
--- a/web/src/lib/seedance-video.ts
+++ b/web/src/lib/seedance-video.ts
@@ -1,4 +1,6 @@
 import type { AiConfig } from "@/stores/use-config-store";
+import type { ReferenceImage } from "@/types/image";
+import type { ReferenceAudio, ReferenceVideo } from "@/types/media";
 
 export const SEEDANCE_REFERENCE_LIMITS = {
     images: 9,
@@ -123,3 +125,44 @@ export function boolConfig(value: string | undefined, fallback: boolean) {
     if (value === "false") return false;
     return fallback;
 }
+
+export function seedanceReferenceLabel(kind: "image" | "video" | "audio", index: number) {
+    if (kind === "image") return `图片${index + 1}`;
+    if (kind === "video") return `视频${index + 1}`;
+    return `音频${index + 1}`;
+}
+
+export function buildSeedancePromptText(prompt: string, images: ReferenceImage[], videos: ReferenceVideo[], audios: ReferenceAudio[]) {
+    const labels = [
+        ...images.map((_, index) => seedanceReferenceLabel("image", index)),
+        ...videos.map((_, index) => seedanceReferenceLabel("video", index)),
+        ...audios.map((_, index) => seedanceReferenceLabel("audio", index)),
+    ];
+    const text = prompt.trim();
+    if (!labels.length) return text;
+    return `参考素材编号：${labels.join("、")}。请按这些编号理解提示词中的图片、视频和音频引用。\n\n${text}`;
+}
+
+export function seedanceVideoReferenceError(videos: ReferenceVideo[]) {
+    let totalDurationMs = 0;
+    for (let index = 0; index < videos.length; index += 1) {
+        const video = videos[index];
+        const label = seedanceReferenceLabel("video", index);
+        if (video.bytes && video.bytes > SEEDANCE_REFERENCE_LIMITS.videoMaxBytes) return `${label} 超过 50MB，请压缩后再上传`;
+        if (video.durationMs) {
+            if (video.durationMs < 2000 || video.durationMs > 15000) return `${label} 时长需要在 2-15 秒之间`;
+            totalDurationMs += video.durationMs;
+        }
+        if (video.width && video.height) {
+            if (video.width < 300 || video.width > 6000 || video.height < 300 || video.height > 6000) return `${label} 宽高需要在 300-6000px 之间`;
+            const ratio = video.width / video.height;
+            if (ratio < 0.4 || ratio > 2.5) return `${label} 宽高比需要在 0.4-2.5 之间`;
+            const pixels = video.width * video.height;
+            if (pixels < 640 * 640 || pixels > 2206 * 946) return `${label} 像素总量不符合 Seedance 要求，请转成 480p/720p/1080p 后再上传`;
+        }
+    }
+    if (totalDurationMs > 15000) return "Seedance 参考视频总时长不能超过 15 秒";
+    return "";
+}
+
+export const seedanceVideoReferenceHint = "参考视频需为 mp4/mov，H.264/H.265，FPS 24-60；含真人人脸素材请使用火山授权 asset:// 素材。";
diff --git a/web/src/services/api/video.ts b/web/src/services/api/video.ts
index 37522f8..ef02431 100644
--- a/web/src/services/api/video.ts
+++ b/web/src/services/api/video.ts
@@ -3,7 +3,7 @@ import axios from "axios";
 import { dataUrlToFile } from "@/lib/image-utils";
 import { getMediaBlob, uploadMediaFile, type UploadedFile } from "@/services/file-storage";
 import { imageToDataUrl } from "@/services/image-storage";
-import { boolConfig, isSeedanceVideoConfig, normalizeSeedanceDuration, normalizeSeedanceRatio, normalizeSeedanceResolution, SEEDANCE_REFERENCE_LIMITS } from "@/lib/seedance-video";
+import { boolConfig, buildSeedancePromptText, isSeedanceVideoConfig, normalizeSeedanceDuration, normalizeSeedanceRatio, normalizeSeedanceResolution, seedanceVideoReferenceError, SEEDANCE_REFERENCE_LIMITS } from "@/lib/seedance-video";
 import { buildApiUrl, type AiConfig } from "@/stores/use-config-store";
 import { useUserStore } from "@/stores/use-user-store";
 import type { ReferenceImage } from "@/types/image";
@@ -130,6 +130,8 @@ async function requestSeedanceGeneration(config: AiConfig, model: string, prompt
 }
 
 function assertSeedanceVideoReferences(videoReferences: ReferenceVideo[]) {
+    const error = seedanceVideoReferenceError(videoReferences);
+    if (error) throw new Error(error);
     let total = 0;
     for (const video of videoReferences) {
         if (!video.durationMs) continue;
@@ -156,7 +158,7 @@ function seedanceApiUrl(config: AiConfig, taskId?: string) {
 
 async function buildSeedanceContent(config: AiConfig, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[], audioReferences: ReferenceAudio[]) {
     const content: Array<Record<string, unknown>> = [];
-    const text = prompt.trim();
+    const text = buildSeedancePromptText(prompt, references, videoReferences, audioReferences);
     if (text) content.push({ type: "text", text });
     for (const image of references.slice(0, SEEDANCE_REFERENCE_LIMITS.images)) {
         content.push({ type: "image_url", image_url: { url: await resolveSeedanceImageUrl(config, image) }, role: "reference_image" });
diff --git a/web/src/types/media.ts b/web/src/types/media.ts
index 84d4e2d..dd1454c 100644
--- a/web/src/types/media.ts
+++ b/web/src/types/media.ts
@@ -4,6 +4,9 @@ export type ReferenceVideo = {
     type: string;
     url: string;
     storageKey?: string;
+    bytes?: number;
+    width?: number;
+    height?: number;
     durationMs?: number;
 };
 

From 6d185a6d935b5b8ddfff2dab78e856475f6381eb Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 22:11:36 +0800
Subject: [PATCH 11/14] fix: persist docker sqlite settings path

---
 .env.example          |  1 +
 config/config.go      | 30 ++++++++++++++++++++++++++++++
 config/config_test.go | 33 +++++++++++++++++++++++++++++++++
 docs/pending-test.md  |  1 +
 4 files changed, 65 insertions(+)
 create mode 100644 config/config_test.go

diff --git a/.env.example b/.env.example
index 6b7ae19..cf0319d 100644
--- a/.env.example
+++ b/.env.example
@@ -19,6 +19,7 @@ PORT=8080
 # 数据库配置，默认使用本地 SQLite。
 STORAGE_DRIVER=sqlite
 # sqlite: DATABASE_DSN=data/infinite-canvas.db
+# Docker 部署时建议使用绝对路径，避免工作目录变化后写入临时库：DATABASE_DSN=/app/data/infinite-canvas.db
 # mysql: DATABASE_DSN=user:password@tcp(127.0.0.1:3306)/infinite_canvas?parseTime=true
 # postgres: DATABASE_DSN=postgres://user:password@127.0.0.1:5432/infinite_canvas?sslmode=disable
 DATABASE_DSN=data/infinite-canvas.db
diff --git a/config/config.go b/config/config.go
index 95976b0..a4ed42c 100644
--- a/config/config.go
+++ b/config/config.go
@@ -3,6 +3,8 @@ package config
 import (
 	"crypto/rand"
 	"encoding/base64"
+	"os"
+	"path/filepath"
 	"strings"
 
 	"github.com/caarlos0/env/v11"
@@ -30,6 +32,7 @@ func Load() error {
 	if err := env.Parse(&Cfg); err != nil {
 		return err
 	}
+	normalizeDockerSQLiteDSN("/app/data")
 	if strings.TrimSpace(Cfg.JWTSecret) == "" || Cfg.JWTSecret == "infinite-canvas" {
 		secret, err := randomSecret()
 		if err != nil {
@@ -40,6 +43,33 @@ func Load() error {
 	return nil
 }
 
+func normalizeDockerSQLiteDSN(appDataDir string) {
+	driver := strings.ToLower(strings.TrimSpace(Cfg.StorageDriver))
+	if driver != "" && driver != "sqlite" {
+		return
+	}
+	dsn := strings.TrimSpace(Cfg.DatabaseDSN)
+	if dsn == "" || dsn == ":memory:" || strings.HasPrefix(dsn, "file:") {
+		return
+	}
+	pathPart, suffix := dsn, ""
+	if index := strings.Index(dsn, "?"); index >= 0 {
+		pathPart = dsn[:index]
+		suffix = dsn[index:]
+	}
+	if filepath.IsAbs(pathPart) {
+		return
+	}
+	slashPath := filepath.ToSlash(pathPart)
+	if slashPath != "data" && !strings.HasPrefix(slashPath, "data/") {
+		return
+	}
+	if _, err := os.Stat(appDataDir); err != nil {
+		return
+	}
+	Cfg.DatabaseDSN = filepath.Join(filepath.Dir(appDataDir), filepath.FromSlash(slashPath)) + suffix
+}
+
 func randomSecret() (string, error) {
 	buf := make([]byte, 32)
 	if _, err := rand.Read(buf); err != nil {
diff --git a/config/config_test.go b/config/config_test.go
new file mode 100644
index 0000000..2f18680
--- /dev/null
+++ b/config/config_test.go
@@ -0,0 +1,33 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestNormalizeDockerSQLiteDSNUsesMountedDataDir(t *testing.T) {
+	root := t.TempDir()
+	appDataDir := filepath.Join(root, "data")
+	if err := os.MkdirAll(appDataDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	Cfg = Config{StorageDriver: "sqlite", DatabaseDSN: "data/infinite-canvas.db?_pragma=busy_timeout(5000)"}
+
+	normalizeDockerSQLiteDSN(appDataDir)
+
+	want := filepath.Join(root, "data", "infinite-canvas.db") + "?_pragma=busy_timeout(5000)"
+	if Cfg.DatabaseDSN != want {
+		t.Fatalf("DatabaseDSN = %q, want %q", Cfg.DatabaseDSN, want)
+	}
+}
+
+func TestNormalizeDockerSQLiteDSNLeavesLocalPathWithoutMountedDataDir(t *testing.T) {
+	Cfg = Config{StorageDriver: "sqlite", DatabaseDSN: "data/infinite-canvas.db"}
+
+	normalizeDockerSQLiteDSN(filepath.Join(t.TempDir(), "missing-data"))
+
+	if Cfg.DatabaseDSN != "data/infinite-canvas.db" {
+		t.Fatalf("DatabaseDSN = %q, want relative local path", Cfg.DatabaseDSN)
+	}
+}
diff --git a/docs/pending-test.md b/docs/pending-test.md
index 24a039d..fc2eb2b 100644
--- a/docs/pending-test.md
+++ b/docs/pending-test.md
@@ -1,5 +1,6 @@
 # 待测试
 
+- Docker 部署时，`DATABASE_DSN=data/infinite-canvas.db` 会在存在 `/app/data` 挂载目录时自动归一到 `/app/data/infinite-canvas.db`，需要验证后台模型配置不会再因为工作目录变为 `/app/web` 而读到空库。
 - Seedance 参考素材失败原因排查：后端会把火山上游错误摘要返回给前端；`/video` 和画布视频生成会按 `图片1/视频1/音频1` 自动编号参考素材，并在实际请求提示词中注入编号说明；参考视频会在请求前校验大小、时长、宽高、宽高比和像素总量。
 - 画布项目导出改为下载 `.zip` 压缩包，包内包含 `projects.json` 和当前画布引用到的本地图片、视频文件，避免只导出 JSON 时丢失媒体内容。
 - 画布库支持多选后一键导出多个画布项目，导出的压缩包可一次恢复多个项目。

From 76461e6dbab4129b1ddfcef018b5c845eddd5c50 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Wed, 27 May 2026 22:34:46 +0800
Subject: [PATCH 12/14] fix: clarify seedance reference video failures

---
 docs/pending-test.md            |  1 +
 handler/ai.go                   | 11 +++++++++++
 handler/ai_test.go              |  7 +++++++
 handler/media_reference.go      | 20 +++++++++++++++++++-
 handler/media_reference_test.go | 18 +++++++++++++++++-
 router/router.go                |  3 +++
 6 files changed, 58 insertions(+), 2 deletions(-)

diff --git a/docs/pending-test.md b/docs/pending-test.md
index fc2eb2b..bdb8a9b 100644
--- a/docs/pending-test.md
+++ b/docs/pending-test.md
@@ -1,6 +1,7 @@
 # 待测试
 
 - Docker 部署时，`DATABASE_DSN=data/infinite-canvas.db` 会在存在 `/app/data` 挂载目录时自动归一到 `/app/data/infinite-canvas.db`，需要验证后台模型配置不会再因为工作目录变为 `/app/web` 而读到空库。
+- Seedance 参考视频被火山判定包含真人或隐私信息时，前端错误摘要会提示改用不含真人的视频、官方允许的模型产物或已授权的 `asset://` 素材；参考素材上传目录改为跟随 SQLite 数据目录，并补充公开素材的 HEAD 访问。
 - Seedance 参考素材失败原因排查：后端会把火山上游错误摘要返回给前端；`/video` 和画布视频生成会按 `图片1/视频1/音频1` 自动编号参考素材，并在实际请求提示词中注入编号说明；参考视频会在请求前校验大小、时长、宽高、宽高比和像素总量。
 - 画布项目导出改为下载 `.zip` 压缩包，包内包含 `projects.json` 和当前画布引用到的本地图片、视频文件，避免只导出 JSON 时丢失媒体内容。
 - 画布库支持多选后一键导出多个画布项目，导出的压缩包可一次恢复多个项目。
diff --git a/handler/ai.go b/handler/ai.go
index c569c0a..26d5d7d 100644
--- a/handler/ai.go
+++ b/handler/ai.go
@@ -263,6 +263,9 @@ func aiUpstreamErrorDetail(body []byte) string {
 	}
 	if err := json.Unmarshal(body, &payload); err == nil {
 		if payload.Error.Message != "" {
+			if detail := friendlyUpstreamError(payload.Error.Code, payload.Error.Message); detail != "" {
+				return safeUpstreamText(detail)
+			}
 			if payload.Error.Code != "" {
 				return safeUpstreamText(payload.Error.Code + " " + payload.Error.Message)
 			}
@@ -278,6 +281,14 @@ func aiUpstreamErrorDetail(body []byte) string {
 	return safeUpstreamText(text)
 }
 
+func friendlyUpstreamError(code string, message string) string {
+	lowerCode := strings.ToLower(strings.TrimSpace(code))
+	if strings.Contains(lowerCode, "inputvideosensitivecontentdetected") || strings.Contains(lowerCode, "privacyinformation") {
+		return strings.TrimSpace(code + " 参考视频疑似包含真人或隐私信息，火山方舟拒绝使用普通 URL 作为真人视频参考；请改用不含真人的视频、官方允许的模型产物，或已授权的 asset:// 素材。原始错误：" + message)
+	}
+	return ""
+}
+
 func safeUpstreamText(text string) string {
 	text = strings.Join(strings.Fields(strings.TrimSpace(text)), " ")
 	runes := []rune(text)
diff --git a/handler/ai_test.go b/handler/ai_test.go
index 2c4d8d4..2eaa897 100644
--- a/handler/ai_test.go
+++ b/handler/ai_test.go
@@ -12,6 +12,13 @@ func TestAIUpstreamErrorDetail(t *testing.T) {
 	}
 }
 
+func TestAIUpstreamErrorDetailExplainsSensitiveVideo(t *testing.T) {
+	got := aiUpstreamErrorDetail([]byte(`{"error":{"code":"InputVideoSensitiveContentDetected.PrivacyInformation","message":"The request failed because the input video may contain real person."}}`))
+	if !strings.Contains(got, "参考视频疑似包含真人") || !strings.Contains(got, "asset://") {
+		t.Fatalf("detail = %q", got)
+	}
+}
+
 func TestSafeUpstreamTextTruncates(t *testing.T) {
 	got := safeUpstreamText(strings.Repeat("错", 320))
 	if len([]rune(got)) != 303 {
diff --git a/handler/media_reference.go b/handler/media_reference.go
index b528588..94d64b7 100644
--- a/handler/media_reference.go
+++ b/handler/media_reference.go
@@ -117,7 +117,25 @@ func ReferenceMedia(w http.ResponseWriter, r *http.Request, id string) {
 }
 
 func referenceMediaDir() string {
-	return filepath.Join("data", "reference-media")
+	return filepath.Join(referenceDataDir(), "reference-media")
+}
+
+func referenceDataDir() string {
+	driver := strings.ToLower(strings.TrimSpace(config.Cfg.StorageDriver))
+	dsn := strings.TrimSpace(config.Cfg.DatabaseDSN)
+	if (driver == "" || driver == "sqlite") && dsn != "" && dsn != ":memory:" && !strings.HasPrefix(dsn, "file:") {
+		pathPart := dsn
+		if index := strings.Index(dsn, "?"); index >= 0 {
+			pathPart = dsn[:index]
+		}
+		if filepath.IsAbs(pathPart) {
+			return filepath.Dir(pathPart)
+		}
+	}
+	if _, err := os.Stat("/app/data"); err == nil {
+		return "/app/data"
+	}
+	return "data"
 }
 
 func normalizeReferenceMediaType(contentType string, ext string) (string, string, bool) {
diff --git a/handler/media_reference_test.go b/handler/media_reference_test.go
index 95cd7fe..826e045 100644
--- a/handler/media_reference_test.go
+++ b/handler/media_reference_test.go
@@ -1,6 +1,11 @@
 package handler
 
-import "testing"
+import (
+	"path/filepath"
+	"testing"
+
+	"github.com/basketikun/infinite-canvas/config"
+)
 
 func TestNormalizeReferenceMediaTypeSupportsAudio(t *testing.T) {
 	tests := []struct {
@@ -38,3 +43,14 @@ func TestReferenceMediaTypeMaxBytes(t *testing.T) {
 		t.Fatalf("image max bytes = %d, want %d", got, referenceImageMaxBytes)
 	}
 }
+
+func TestReferenceMediaDirUsesAbsoluteSQLiteDataDir(t *testing.T) {
+	previous := config.Cfg
+	t.Cleanup(func() { config.Cfg = previous })
+	root := t.TempDir()
+	config.Cfg = config.Config{StorageDriver: "sqlite", DatabaseDSN: filepath.Join(root, "infinite-canvas.db")}
+
+	if got := referenceMediaDir(); got != filepath.Join(root, "reference-media") {
+		t.Fatalf("referenceMediaDir = %q", got)
+	}
+}
diff --git a/router/router.go b/router/router.go
index 81525d5..6493903 100644
--- a/router/router.go
+++ b/router/router.go
@@ -25,6 +25,9 @@ func New() *gin.Engine {
 	api.GET("/media/references/:id", func(c *gin.Context) {
 		handler.ReferenceMedia(c.Writer, c.Request, c.Param("id"))
 	})
+	api.HEAD("/media/references/:id", func(c *gin.Context) {
+		handler.ReferenceMedia(c.Writer, c.Request, c.Param("id"))
+	})
 	v1 := api.Group("/v1", middleware.UserAuth)
 	v1.POST("/images/generations", gin.WrapF(handler.AIImagesGenerations))
 	v1.POST("/images/edits", gin.WrapF(handler.AIImagesEdits))

From 923e21c58124502adc09ad6d4a0570125c1c92e4 Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Thu, 28 May 2026 14:28:53 +0800
Subject: [PATCH 13/14] fix: normalize gpt image request size

---
 docs/pending-test.md                        |  1 +
 web/src/components/image-settings-panel.tsx |  7 +-
 web/src/services/api/image.ts               | 78 +++++++++++++++------
 3 files changed, 62 insertions(+), 24 deletions(-)

diff --git a/docs/pending-test.md b/docs/pending-test.md
index bdb8a9b..2db3ddf 100644
--- a/docs/pending-test.md
+++ b/docs/pending-test.md
@@ -1,5 +1,6 @@
 # 待测试
 
+- GPT Image 生图请求会在前端把 `9:16`、`16:9` 等比例转换成合法 `WIDTHxHEIGHT` 尺寸，并在非法尺寸时直接显示中文错误，避免上游返回 `invalid_value Invalid size`。
 - Docker 部署时，`DATABASE_DSN=data/infinite-canvas.db` 会在存在 `/app/data` 挂载目录时自动归一到 `/app/data/infinite-canvas.db`，需要验证后台模型配置不会再因为工作目录变为 `/app/web` 而读到空库。
 - Seedance 参考视频被火山判定包含真人或隐私信息时，前端错误摘要会提示改用不含真人的视频、官方允许的模型产物或已授权的 `asset://` 素材；参考素材上传目录改为跟随 SQLite 数据目录，并补充公开素材的 HEAD 访问。
 - Seedance 参考素材失败原因排查：后端会把火山上游错误摘要返回给前端；`/video` 和画布视频生成会按 `图片1/视频1/音频1` 自动编号参考素材，并在实际请求提示词中注入编号说明；参考视频会在请求前校验大小、时长、宽高、宽高比和像素总量。
diff --git a/web/src/components/image-settings-panel.tsx b/web/src/components/image-settings-panel.tsx
index 66025ce..61095b7 100644
--- a/web/src/components/image-settings-panel.tsx
+++ b/web/src/components/image-settings-panel.tsx
@@ -17,9 +17,10 @@ const aspectOptions = [
     { value: "1:1", label: "1:1", width: 1024, height: 1024, icon: "square" },
     { value: "3:2", label: "3:2", width: 1536, height: 1024, icon: "landscape" },
     { value: "2:3", label: "2:3", width: 1024, height: 1536, icon: "portrait" },
-    { value: "4:3", label: "4:3", width: 1344, height: 1024, icon: "landscape" },
-    { value: "3:4", label: "3:4", width: 1024, height: 1344, icon: "portrait" },
-    { value: "9:16", label: "9:16", width: 1024, height: 1792, icon: "portrait" },
+    { value: "4:3", label: "4:3", width: 1360, height: 1024, icon: "landscape" },
+    { value: "3:4", label: "3:4", width: 1024, height: 1360, icon: "portrait" },
+    { value: "16:9", label: "16:9", width: 1824, height: 1024, icon: "landscape" },
+    { value: "9:16", label: "9:16", width: 1024, height: 1824, icon: "portrait" },
     { value: "1:1-2k", label: "1:1(2k)", size: "2048x2048", width: 2048, height: 2048, icon: "square" },
     { value: "16:9-2k", label: "16:9(2k)", size: "2048x1152", width: 2048, height: 1152, icon: "landscape" },
     { value: "9:16-2k", label: "9:16(2k)", size: "1152x2048", width: 1152, height: 2048, icon: "portrait" },
diff --git a/web/src/services/api/image.ts b/web/src/services/api/image.ts
index dea317e..87319e2 100644
--- a/web/src/services/api/image.ts
+++ b/web/src/services/api/image.ts
@@ -31,6 +31,12 @@ const QUALITY_ALIASES: Record<string, string> = {
     "2k": "medium",
     "4k": "high",
 };
+const DEFAULT_IMAGE_SHORT_SIDE = 1024;
+const IMAGE_SIZE_STEP = 16;
+const IMAGE_MIN_PIXELS = 655360;
+const IMAGE_MAX_PIXELS = 8294400;
+const IMAGE_MAX_EDGE = 3840;
+const IMAGE_MAX_RATIO = 3;
 
 function normalizeQuality(quality: string) {
     const value = quality.trim().toLowerCase();
@@ -38,36 +44,66 @@ function normalizeQuality(quality: string) {
     return QUALITY_BASE[normalized] ? normalized : undefined;
 }
 
-/** Map "quality + ratio" to an explicit pixel dimension like "3840x2160". Returns undefined when quality is auto. */
-function resolveSize(quality: string, ratio: string): string | undefined {
-    const basePixels = QUALITY_BASE[quality];
-    if (!basePixels || ratio === "auto" || !ratio) return undefined;
+/** Map "quality + ratio" to an explicit pixel dimension like "3840x2160". */
+function resolveSize(quality: string | undefined, ratio: string): string {
+    const parsedRatio = parseImageRatio(ratio);
+    const basePixels = quality ? QUALITY_BASE[quality] : undefined;
+    const isLandscape = parsedRatio.width >= parsedRatio.height;
+    const longRatio = isLandscape ? parsedRatio.width / parsedRatio.height : parsedRatio.height / parsedRatio.width;
+    let longSide: number;
+    let shortSide: number;
 
-    const parts = ratio.split(":");
-    if (parts.length !== 2) return undefined;
-    const w = Number(parts[0]);
-    const h = Number(parts[1]);
-    if (!w || !h) return undefined;
-
-    const targetPixels = basePixels * basePixels;
-    const isLandscape = w >= h;
-    const longRatio = isLandscape ? w / h : h / w;
-
-    const longSideRaw = Math.sqrt(targetPixels * longRatio);
-    const longSide = Math.floor(longSideRaw / 16) * 16;
-    const shortSide = Math.round((longSide / longRatio) / 16) * 16;
+    if (basePixels) {
+        const targetPixels = basePixels * basePixels;
+        const longSideRaw = Math.sqrt(targetPixels * longRatio);
+        longSide = Math.floor(longSideRaw / IMAGE_SIZE_STEP) * IMAGE_SIZE_STEP;
+        shortSide = Math.round(longSide / longRatio / IMAGE_SIZE_STEP) * IMAGE_SIZE_STEP;
+    } else {
+        shortSide = DEFAULT_IMAGE_SHORT_SIDE;
+        longSide = Math.round((shortSide * longRatio) / IMAGE_SIZE_STEP) * IMAGE_SIZE_STEP;
+    }
 
     const width = isLandscape ? longSide : shortSide;
     const height = isLandscape ? shortSide : longSide;
-
+    validateImageSize(width, height);
     return `${width}x${height}`;
 }
 
+function parseImageRatio(value: string) {
+    const parts = value.split(":");
+    if (parts.length !== 2) throw new Error("图像尺寸格式不支持，请使用 auto、9:16 或 1024x1024");
+    const w = Number(parts[0]);
+    const h = Number(parts[1]);
+    if (!Number.isFinite(w) || !Number.isFinite(h) || w <= 0 || h <= 0) throw new Error("图像比例必须是正数，例如 9:16");
+    if (Math.max(w, h) / Math.min(w, h) > IMAGE_MAX_RATIO) throw new Error("图像宽高比不能超过 3:1，请调整尺寸");
+    return { width: w, height: h };
+}
+
+function parseImageDimensions(value: string) {
+    const match = value.match(/^(\d+)x(\d+)$/i);
+    if (!match) return null;
+    return { width: Number(match[1]), height: Number(match[2]) };
+}
+
+function validateImageSize(width: number, height: number) {
+    if (!Number.isInteger(width) || !Number.isInteger(height) || width <= 0 || height <= 0) throw new Error("图像尺寸必须是正整数，例如 1024x1024");
+    if (width % IMAGE_SIZE_STEP !== 0 || height % IMAGE_SIZE_STEP !== 0) throw new Error("图像尺寸的宽高必须是 16 的倍数，请调整尺寸");
+    if (Math.max(width, height) > IMAGE_MAX_EDGE) throw new Error("图像尺寸最长边不能超过 3840px，请调整尺寸");
+    if (Math.max(width, height) / Math.min(width, height) > IMAGE_MAX_RATIO) throw new Error("图像宽高比不能超过 3:1，请调整尺寸");
+    const pixels = width * height;
+    if (pixels < IMAGE_MIN_PIXELS || pixels > IMAGE_MAX_PIXELS) throw new Error("图像总像素需在 655360 到 8294400 之间，请调整尺寸");
+}
+
 function resolveRequestSize(quality: string | undefined, size: string) {
     const value = size.trim();
-    if (!value || value === "auto") return undefined;
-    if (/^\d+x\d+$/.test(value)) return value;
-    return (quality && resolveSize(quality, value)) || value;
+    if (!value || value.toLowerCase() === "auto") return undefined;
+    const dimensions = parseImageDimensions(value);
+    if (dimensions) {
+        validateImageSize(dimensions.width, dimensions.height);
+        return `${dimensions.width}x${dimensions.height}`;
+    }
+    if (value.includes(":")) return resolveSize(quality, value);
+    throw new Error("图像尺寸格式不支持，请使用 auto、9:16 或 1024x1024");
 }
 
 function resolveImageDataUrl(item: Record<string, unknown>) {

From c3b1ba2943802819c4696ed12ead384a1821231d Mon Sep 17 00:00:00 2001
From: stupid-h4er <stupid-h4er@users.noreply.github.com>
Date: Sun, 31 May 2026 00:54:54 +0800
Subject: [PATCH 14/14] feat: label image references in generation

---
 docs/pending-test.md                          |  1 +
 .../components/canvas-assistant-panel.tsx     | 22 +++++++++++-----
 .../components/canvas-config-node-panel.tsx   |  3 ++-
 web/src/app/(user)/image/page.tsx             | 25 +++++++++++++++++--
 web/src/lib/image-reference-prompt.ts         | 12 +++++++++
 web/src/services/api/image.ts                 |  4 ++-
 6 files changed, 57 insertions(+), 10 deletions(-)
 create mode 100644 web/src/lib/image-reference-prompt.ts

diff --git a/docs/pending-test.md b/docs/pending-test.md
index 2db3ddf..94a9502 100644
--- a/docs/pending-test.md
+++ b/docs/pending-test.md
@@ -1,5 +1,6 @@
 # 待测试
 
+- 生图工作台和画布生图会把参考图按当前顺序显示为 `图片1`、`图片2` 等编号，并在图生图请求的实际提示词中注入编号说明；需要验证 `/image` 参考图排序、画布配置节点输入顺序和画布助手参考图编号一致。
 - GPT Image 生图请求会在前端把 `9:16`、`16:9` 等比例转换成合法 `WIDTHxHEIGHT` 尺寸，并在非法尺寸时直接显示中文错误，避免上游返回 `invalid_value Invalid size`。
 - Docker 部署时，`DATABASE_DSN=data/infinite-canvas.db` 会在存在 `/app/data` 挂载目录时自动归一到 `/app/data/infinite-canvas.db`，需要验证后台模型配置不会再因为工作目录变为 `/app/web` 而读到空库。
 - Seedance 参考视频被火山判定包含真人或隐私信息时，前端错误摘要会提示改用不含真人的视频、官方允许的模型产物或已授权的 `asset://` 素材；参考素材上传目录改为跟随 SQLite 数据目录，并补充公开素材的 HEAD 访问。
diff --git a/web/src/app/(user)/canvas/components/canvas-assistant-panel.tsx b/web/src/app/(user)/canvas/components/canvas-assistant-panel.tsx
index 8860bee..0b10338 100644
--- a/web/src/app/(user)/canvas/components/canvas-assistant-panel.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-assistant-panel.tsx
@@ -16,6 +16,7 @@ import { requestEdit, requestGeneration, requestImageQuestion, type ChatCompleti
 import { imageToDataUrl, uploadImage } from "@/services/image-storage";
 import { useAssetStore } from "@/stores/use-asset-store";
 import { useThemeStore } from "@/stores/use-theme-store";
+import { imageReferenceLabel } from "@/lib/image-reference-prompt";
 import type { ReferenceImage } from "@/types/image";
 import { DiaTextReveal } from "@/components/ui/dia-text-reveal";
 import { CanvasImageSettingsPopover } from "./canvas-image-settings-popover";
@@ -398,8 +399,8 @@ function AssistantComposer({
         <div className="px-2 pb-2" onWheelCapture={(event) => event.stopPropagation()}>
             {references.length ? (
                 <div className="thin-scrollbar mb-1.5 flex max-w-full gap-1.5 overflow-x-auto px-1 pb-1">
-                    {references.map((item) => (
-                        <AssistantReferenceChip key={item.id} item={item} onRemove={() => onRemoveReference(item.id)} />
+                    {references.map((item, index) => (
+                        <AssistantReferenceChip key={item.id} item={item} label={assistantImageReferenceLabel(references, index)} onRemove={() => onRemoveReference(item.id)} />
                     ))}
                 </div>
             ) : null}
@@ -584,20 +585,23 @@ function AssistantHistory({
 function MessageReferences({ message }: { message: CanvasAssistantMessage }) {
     return (
         <div className={cn("flex max-w-[88%] flex-wrap gap-2", message.role === "user" ? "justify-end" : "justify-start")}>
-            {message.references?.map((item) => (
-                <AssistantReferenceChip key={item.id} item={item} />
+            {message.references?.map((item, index, references) => (
+                <AssistantReferenceChip key={item.id} item={item} label={assistantImageReferenceLabel(references, index)} />
             ))}
         </div>
     );
 }
 
-function AssistantReferenceChip({ item, onRemove }: { item: CanvasAssistantReference; onRemove?: () => void }) {
+function AssistantReferenceChip({ item, label, onRemove }: { item: CanvasAssistantReference; label?: string; onRemove?: () => void }) {
     const theme = canvasThemes[useThemeStore((state) => state.theme)];
     const text = (item.text || item.title).replace(/\s+/g, " ").trim().slice(0, 1) || "文";
     return (
         <div className="group/chip relative inline-flex h-8 max-w-[150px] shrink-0 items-center gap-1.5 rounded-lg text-sm" style={{ color: theme.node.text }}>
             {item.dataUrl ? (
-                <img src={item.dataUrl} alt="" className="size-8 rounded-lg object-cover" />
+                <span className="relative block size-8 shrink-0">
+                    <img src={item.dataUrl} alt="" className="size-8 rounded-lg object-cover" />
+                    {label ? <span className="absolute left-0.5 top-0.5 rounded bg-black/60 px-1 py-0.5 text-[8px] font-medium leading-none text-white">{label}</span> : null}
+                </span>
             ) : (
                 <span className="grid size-8 place-items-center rounded-lg border text-sm font-medium" style={{ background: theme.node.panel, borderColor: theme.node.activeStroke }}>
                     {text}
@@ -618,6 +622,12 @@ function AssistantReferenceChip({ item, onRemove }: { item: CanvasAssistantRefer
     );
 }
 
+function assistantImageReferenceLabel(references: CanvasAssistantReference[], index: number) {
+    if (!references[index]?.dataUrl) return undefined;
+    const imageIndex = references.slice(0, index + 1).filter((item) => item.dataUrl).length - 1;
+    return imageIndex >= 0 ? imageReferenceLabel(imageIndex) : undefined;
+}
+
 function nodeToReference(node: CanvasNodeData): CanvasAssistantReference | null {
     if (node.type === CanvasNodeType.Image && node.metadata?.content) {
         return { id: node.id, type: node.type, title: node.title, dataUrl: node.metadata.content, storageKey: node.metadata.storageKey };
diff --git a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
index f910ed7..4116989 100644
--- a/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
+++ b/web/src/app/(user)/canvas/components/canvas-config-node-panel.tsx
@@ -9,6 +9,7 @@ import { ModelPicker } from "@/components/model-picker";
 import { defaultConfig, useConfigStore, useEffectiveConfig, type AiConfig } from "@/stores/use-config-store";
 import { CreditSymbol, requestCreditCost } from "@/constant/credits";
 import { canvasThemes } from "@/lib/canvas-theme";
+import { imageReferenceLabel } from "@/lib/image-reference-prompt";
 import { seedanceReferenceLabel } from "@/lib/seedance-video";
 import { useThemeStore } from "@/stores/use-theme-store";
 import { CanvasImageSettingsPopover } from "./canvas-image-settings-popover";
@@ -304,7 +305,7 @@ function ImageSortCard({
         <div className="w-24 shrink-0 overflow-hidden rounded-lg border" style={{ background: theme.node.fill, borderColor: theme.node.stroke }}>
             <div className="relative">
                 <img src={input.image.dataUrl} alt={input.title} className="aspect-square w-full object-cover" />
-                <span className="absolute left-1 top-1 rounded bg-black/50 px-1 py-0.5 text-[9px] font-medium text-white">{seedanceReferenceLabel("image", imageIndex)}</span>
+                <span className="absolute left-1 top-1 rounded bg-black/50 px-1 py-0.5 text-[9px] font-medium text-white">{imageReferenceLabel(imageIndex)}</span>
                 <HorizontalOrderButtons index={imageIndex} total={imageTotal} onMove={(offset) => onMove(input, offset)} />
             </div>
         </div>
diff --git a/web/src/app/(user)/image/page.tsx b/web/src/app/(user)/image/page.tsx
index 34be005..44e06f5 100644
--- a/web/src/app/(user)/image/page.tsx
+++ b/web/src/app/(user)/image/page.tsx
@@ -1,6 +1,6 @@
 "use client";
 
-import { BookOpen, CheckSquare, ClipboardPaste, Download, FolderPlus, History, ImagePlus, LoaderCircle, PenLine, Plus, SlidersHorizontal, Sparkles, Trash2, Upload } from "lucide-react";
+import { ArrowLeft, ArrowRight, BookOpen, CheckSquare, ClipboardPaste, Download, FolderPlus, History, ImagePlus, LoaderCircle, PenLine, Plus, SlidersHorizontal, Sparkles, Trash2, Upload } from "lucide-react";
 import { useEffect, useRef, useState } from "react";
 import { App, Button, Checkbox, Drawer, Empty, Image, Input, Modal, Tag, Typography } from "antd";
 import localforage from "localforage";
@@ -11,6 +11,7 @@ import { ModelPicker } from "@/components/model-picker";
 import { PromptSelectDialog } from "@/components/prompts/prompt-select-dialog";
 import { AssetPickerModal, type InsertAssetPayload } from "@/app/(user)/canvas/components/asset-picker-modal";
 import { canvasThemes } from "@/lib/canvas-theme";
+import { imageReferenceLabel } from "@/lib/image-reference-prompt";
 import { useConfigStore, useEffectiveConfig, type AiConfig } from "@/stores/use-config-store";
 import { useThemeStore } from "@/stores/use-theme-store";
 import { nanoid } from "nanoid";
@@ -373,9 +374,11 @@ export default function ImagePage() {
                                         event.currentTarget.scrollLeft += event.deltaY;
                                     }}
                                 >
-                                    {references.map((item) => (
+                                    {references.map((item, index) => (
                                         <div key={item.id} className="group relative size-20 shrink-0 overflow-hidden rounded-md border border-stone-200 dark:border-stone-800">
                                             <img src={item.dataUrl} alt={item.name} className="size-full object-cover" />
+                                            <span className="absolute left-1 top-1 rounded bg-black/60 px-1.5 py-0.5 text-[10px] font-medium text-white">{imageReferenceLabel(index)}</span>
+                                            <ReferenceOrderButtons index={index} total={references.length} onMove={(offset) => setReferences((value) => moveListItem(value, index, offset))} />
                                             <button
                                                 type="button"
                                                 className="absolute right-1 top-1 hidden size-6 items-center justify-center rounded bg-black/60 text-white group-hover:flex"
@@ -742,6 +745,24 @@ function normalizeLogConfig(log: Partial<GenerationLog>): GenerationLogConfig {
     };
 }
 
+function moveListItem<T>(items: T[], index: number, offset: number) {
+    const targetIndex = index + offset;
+    if (targetIndex < 0 || targetIndex >= items.length) return items;
+    const next = [...items];
+    [next[index], next[targetIndex]] = [next[targetIndex], next[index]];
+    return next;
+}
+
+function ReferenceOrderButtons({ index, total, onMove }: { index: number; total: number; onMove: (offset: number) => void }) {
+    if (total <= 1) return null;
+    return (
+        <div className="absolute inset-x-1 bottom-1 flex justify-between">
+            <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !bg-white/85 !p-0 !shadow-sm" icon={<ArrowLeft className="size-3" />} disabled={index <= 0} onClick={() => onMove(-1)} />
+            <Button size="small" className="!h-6 !w-6 !min-w-6 !rounded-full !bg-white/85 !p-0 !shadow-sm" icon={<ArrowRight className="size-3" />} disabled={index >= total - 1} onClick={() => onMove(1)} />
+        </div>
+    );
+}
+
 function buildLog({
     prompt,
     model,
diff --git a/web/src/lib/image-reference-prompt.ts b/web/src/lib/image-reference-prompt.ts
new file mode 100644
index 0000000..3b99ba4
--- /dev/null
+++ b/web/src/lib/image-reference-prompt.ts
@@ -0,0 +1,12 @@
+import type { ReferenceImage } from "@/types/image";
+
+export function imageReferenceLabel(index: number) {
+    return `图片${index + 1}`;
+}
+
+export function buildImageReferencePromptText(prompt: string, references: ReferenceImage[]) {
+    const text = prompt.trim();
+    if (!references.length) return text;
+    const labels = references.map((_, index) => imageReferenceLabel(index));
+    return `参考图片编号：${labels.join("、")}。请按这些编号理解提示词中的图片引用。\n\n${text}`;
+}
diff --git a/web/src/services/api/image.ts b/web/src/services/api/image.ts
index 87319e2..3bc5b8f 100644
--- a/web/src/services/api/image.ts
+++ b/web/src/services/api/image.ts
@@ -4,6 +4,7 @@ import { buildApiUrl, type AiConfig } from "@/stores/use-config-store";
 import { useUserStore } from "@/stores/use-user-store";
 import { nanoid } from "nanoid";
 import { dataUrlToFile } from "@/lib/image-utils";
+import { buildImageReferencePromptText } from "@/lib/image-reference-prompt";
 import { imageToDataUrl } from "@/services/image-storage";
 import type { ReferenceImage } from "@/types/image";
 
@@ -223,9 +224,10 @@ export async function requestEdit(config: AiConfig, prompt: string, references:
     const n = Math.max(1, Math.min(15, Math.floor(Math.abs(Number(config.count)) || 1)));
     const quality = normalizeQuality(config.quality);
     const requestSize = resolveRequestSize(quality, config.size);
+    const requestPrompt = buildImageReferencePromptText(prompt, references);
     const formData = new FormData();
     formData.set("model", config.model);
-    formData.set("prompt", withSystemPrompt(config, prompt));
+    formData.set("prompt", withSystemPrompt(config, requestPrompt));
     formData.set("n", String(n));
     formData.set("response_format", "b64_json");
     if (quality) {