fix(canvas): 视频生成 - 仅使用输入框提示词并保留结构化媒体引用

backend/internal/service/provider.go

@@ -162,8 +162,12 @@ func (s *Service) processCanvasGenerationTask(ctx context.Context, userID string
 	if strings.TrimSpace(input.Prompt) == "" {
 		input.Prompt = fallbackPrompt
 	}
+	if input.Mode == "" && strings.HasPrefix(taskType, "video_") {
+		input.Mode = "video"
+	}
 	promptTemplateOperation := metadataString(input.Metadata, "promptTemplateOperation")
-	if promptTemplateOperation != "" {
+	// 视频节点的最终 Prompt 只取输入框内容，不能被分镜模板替换；图片和文本仍沿用模板能力。
+	if input.Mode != "video" && promptTemplateOperation != "" {
 		values := metadataStringValues(input.Metadata["promptTemplateVariables"])
 		compiled, compileErr := s.compilePrompt(userID, promptTemplateOperation, values)
 		if compileErr != nil {
@@ -174,15 +178,12 @@ func (s *Service) processCanvasGenerationTask(ctx context.Context, userID string
 	if strings.TrimSpace(input.Prompt) == "" {
 		return nil, errors.New("prompt is required")
 	}
-	if input.Mode == "" && strings.HasPrefix(taskType, "video_") {
-		input.Mode = "video"
-	}
 	config, err := s.resolveProviderConfig(input.Config)
 	if err != nil {
 		return nil, err
 	}
 	input.Config = config
-	if (input.Mode == "image" || input.Mode == "video") && input.Metadata != nil {
+	if input.Mode == "image" && input.Metadata != nil {
 		if err := s.applyGenerationStyleProfile(userID, taskProjectID, &input); err != nil {
 			return nil, err
 		}

web/src/components/canvas/canvas-node-generation.ts

@@ -55,13 +55,13 @@ export type NodeGenerationInput = {
     character?: CharacterGenerationReference;
 };
 
-export function buildNodeGenerationContext(nodeId: string, nodes: CanvasNodeData[], connections: CanvasConnection[], prompt: string): NodeGenerationContext {
+export function buildNodeGenerationContext(nodeId: string, nodes: CanvasNodeData[], connections: CanvasConnection[], prompt: string, promptOnly = false): NodeGenerationContext {
     const inputs = buildNodeGenerationInputs(nodeId, nodes, connections);
     const sourceNode = nodes.find((node) => node.id === nodeId);
     const storyboardInputs = getConnectedStoryboardRows(nodeId, nodes, connections);
     const hasExplicitNodeMention = /@\[node:[^\]]+\]/.test(normalizeLegacyNodeMentions(prompt, inputs));
     if ((sourceNode?.type === CanvasNodeType.Config && Boolean(sourceNode.metadata?.composerContent?.trim())) || hasExplicitNodeMention) {
-        return buildComposerGenerationContext(inputs, prompt, [sourceNode?.metadata?.videoStartFrameNodeId, sourceNode?.metadata?.videoEndFrameNodeId].filter((id): id is string => Boolean(id)));
+        return buildComposerGenerationContext(inputs, prompt, [sourceNode?.metadata?.videoStartFrameNodeId, sourceNode?.metadata?.videoEndFrameNodeId].filter((id): id is string => Boolean(id)), promptOnly);
     }
 
     const isStoryboardMedia = sourceNode?.type === CanvasNodeType.Image || sourceNode?.type === CanvasNodeType.Video;
@@ -77,7 +77,7 @@ export function buildNodeGenerationContext(nodeId: string, nodes: CanvasNodeData
     const referenceAudios = inputs.map((input) => input.audio).filter((audio): audio is ReferenceAudio => Boolean(audio));
 
     return {
-        prompt: upstreamText ? `${basePrompt}\n\n${upstreamText}` : basePrompt,
+        prompt: promptOnly ? prompt : upstreamText ? `${basePrompt}\n\n${upstreamText}` : basePrompt,
         referenceImages,
         referenceVideos,
         referenceAudios,
@@ -109,7 +109,7 @@ function removeTrailingInputBlocks(prompt: string, inputs: NodeGenerationInput[]
     return next;
 }
 
-function buildComposerGenerationContext(inputs: NodeGenerationInput[], prompt: string, videoFrameNodeIds: string[] = []): NodeGenerationContext {
+function buildComposerGenerationContext(inputs: NodeGenerationInput[], prompt: string, videoFrameNodeIds: string[] = [], promptOnly = false): NodeGenerationContext {
     const normalizedPrompt = normalizeLegacyNodeMentions(prompt, inputs);
     const inputByNodeId = new Map(inputs.map((input) => [input.nodeId, input]));
     const selectedInputs: NodeGenerationInput[] = [];
@@ -140,7 +140,7 @@ function buildComposerGenerationContext(inputs: NodeGenerationInput[], prompt: s
     }
 
     nextPrompt += normalizedPrompt.slice(lastIndex);
-    if (textBlocks.length) nextPrompt = `${nextPrompt.trim()}\n\n${textBlocks.join("\n\n")}`;
+    if (textBlocks.length && !promptOnly) nextPrompt = `${nextPrompt.trim()}\n\n${textBlocks.join("\n\n")}`;
     // 首尾帧是结构化生成参数，不受提示词中的 @ 引用筛选影响。
     const selectedNodeIds = new Set(selectedInputs.map((input) => input.nodeId));
     videoFrameNodeIds.forEach((nodeId) => {
@@ -286,7 +286,7 @@ export function buildNodeResponseMessages(context: NodeGenerationContext): AiTex
     ];
 }
 
-export async function hydrateNodeGenerationContext(context: NodeGenerationContext, projectId: string, domainProjectId?: string, mode?: CanvasGenerationMode, includeCharacterVoiceSamples = false) {
+export async function hydrateNodeGenerationContext(context: NodeGenerationContext, projectId: string, domainProjectId?: string, mode?: CanvasGenerationMode, includeCharacterVoiceSamples = false, includeCharacterPrompt = true) {
     const { imageToDataUrl } = await import("@/services/image-storage");
     let referenceImages = await Promise.all(
         context.referenceImages.map(async (image) => {
@@ -365,7 +365,7 @@ export async function hydrateNodeGenerationContext(context: NodeGenerationContex
     const voiceBlocks = mode === "video" ? resolvedCharacterVoices.map(compileResolvedVoicePrompt) : [];
     return {
         ...context,
-        prompt: [context.prompt.trim(), ...characterBlocks, ...voiceBlocks].filter(Boolean).join("\n\n"),
+        prompt: includeCharacterPrompt ? [context.prompt.trim(), ...characterBlocks, ...voiceBlocks].filter(Boolean).join("\n\n") : context.prompt,
         referenceImages,
         referenceAudios,
         resolvedCharacterVersions,

web/src/pages/canvas/use-canvas-generation-executor.ts

@@ -124,13 +124,16 @@ export function useCanvasGenerationExecutor({
             }
 
             let rawGenerationContext: Awaited<ReturnType<typeof hydrateNodeGenerationContext>>;
+            // 视频文本只保留输入框内容；连接的媒体仍作为结构化参考传递。
+            const promptOnly = mode === "video";
             try {
                 rawGenerationContext = await hydrateNodeGenerationContext(
-                    buildNodeGenerationContext(nodeId, nodesRef.current, connectionsRef.current, editingTextNode ? `请根据要求修改以下文本。\n\n原文：\n${sourceTextContent}\n\n修改要求：\n${prompt}` : generationPrompt),
+                    buildNodeGenerationContext(nodeId, nodesRef.current, connectionsRef.current, editingTextNode ? `请根据要求修改以下文本。\n\n原文：\n${sourceTextContent}\n\n修改要求：\n${prompt}` : generationPrompt, promptOnly),
                     projectId,
                     domainProjectId,
                     mode,
                     mode === "video" && supportsVideoReferenceAudio(generationConfig),
+                    !promptOnly,
                 );
             } catch (error) {
                 const errorDetails = generationErrorMessage(error);
@@ -143,10 +146,10 @@ export function useCanvasGenerationExecutor({
                 return;
             }
 
-            const expandedPrompt = expandSkillMentions(rawGenerationContext.prompt, addedSkills);
+            const expandedPrompt = promptOnly ? rawGenerationContext.prompt : expandSkillMentions(rawGenerationContext.prompt, addedSkills);
             let effectivePrompt = expandedPrompt.trim();
             let styleMetadata = {};
-            if (mode === "image" || mode === "video") {
+            if (mode === "image") {
                 try {
                     const styleRuntime = resolveCanvasStyleExecution(nodesRef.current, sourceNode, effectivePrompt, generationConfig, mode);
                     if (styleRuntime) {

web/src/pages/canvas/use-canvas-generation-retry.ts

@@ -86,15 +86,16 @@ export function useCanvasGenerationRetry({ projectId, domainProjectId, addedSkil
             }
             let rawContext: Awaited<ReturnType<typeof hydrateNodeGenerationContext>> | null;
             try {
-                const baseContext = buildNodeGenerationContext(sourceNode.id, nodesRef.current, connectionsRef.current, retryContextPrompt);
-                rawContext = hasSavedImageMetadata && !baseContext.characterReferences.length ? null : await hydrateNodeGenerationContext(baseContext, projectId, domainProjectId, retryMode, retryMode === "video" && supportsVideoReferenceAudio(generationConfig));
+                const promptOnly = retryMode === "video";
+                const baseContext = buildNodeGenerationContext(sourceNode.id, nodesRef.current, connectionsRef.current, retryContextPrompt, promptOnly);
+                rawContext = hasSavedImageMetadata && !baseContext.characterReferences.length ? null : await hydrateNodeGenerationContext(baseContext, projectId, domainProjectId, retryMode, retryMode === "video" && supportsVideoReferenceAudio(generationConfig), !promptOnly);
             } catch (error) {
                 const failure = generationFailureMetadata(error, retryPromptSource);
                 message.error(failure.errorDetails);
                 setNodes((current) => current.map((item) => (item.id === node.id ? { ...item, metadata: { ...item.metadata, status: NODE_STATUS_ERROR, ...failure } } : item)));
                 return;
             }
-            const context = rawContext ? { ...rawContext, prompt: expandSkillMentions(rawContext.prompt, addedSkills) } : null;
+            const context = rawContext ? { ...rawContext, prompt: retryMode === "video" ? rawContext.prompt : expandSkillMentions(rawContext.prompt, addedSkills) } : null;
             const prompt = (context?.characterReferences.length ? context.prompt : savedImageMetadata?.prompt || context?.prompt || "").trim();
             if (!prompt) {
                 message.warning("找不到提示词，无法重试");
@@ -102,9 +103,9 @@ export function useCanvasGenerationRetry({ projectId, domainProjectId, addedSkil
             }
             let mediaPrompt = prompt;
             let styleMetadata = {};
-            if (node.type === CanvasNodeType.Image || node.type === CanvasNodeType.Video) {
+            if (node.type === CanvasNodeType.Image) {
                 try {
-                    const runtime = resolveCanvasStyleExecution(nodesRef.current, sourceNode, prompt, generationConfig, node.type === CanvasNodeType.Image ? "image" : "video");
+                    const runtime = resolveCanvasStyleExecution(nodesRef.current, sourceNode, prompt, generationConfig, "image");
                     if (runtime) {
                         mediaPrompt = runtime.prompt;
                         styleMetadata = { styleProfileJson: runtime.profileJson, styleExecutionPlan: runtime.plan };