Add official Alibaba audio generation and editable director object tracks

2026-09-17 15:43:32 +08:00 · 2026-09-17 15:43:32 +08:00 · cb42f27ee9
commit cb42f27ee9
parent 292eddc5a6
27 changed files with 2188 additions and 125 deletions
--- a/app/api/canvas/audio/models/route.ts
+++ b/app/api/canvas/audio/models/route.ts
@ -0,0 +1,24 @@
+import { NextResponse } from "next/server";
+import { auth } from "@/app/(auth)/auth";
+import { listCanvasAudioModels } from "@/lib/video-canvas/audio-generation";
+import { audioContract } from "@/lib/video-canvas/audio-contract";
+export async function GET() {
+  const session = await auth();
+  if (!session?.user?.id) return NextResponse.json({ error: "未登录" }, { status: 401 });
+  try {
+    const models = await listCanvasAudioModels();
+    return NextResponse.json(
+      {
+        models: models.map(({ model }) => ({
+          id: model.id,
+          modelId: model.modelId,
+          name: model.displayName,
+          ...audioContract(model.modelId),
+        })),
+      },
+      { headers: { "Cache-Control": "no-store" } },
+    );
+  } catch {
+    return NextResponse.json({ error: "音频模型读取失败，请稍后重试" }, { status: 503 });
+  }
+}
--- a/app/api/canvas/generation-tasks/route.ts
+++ b/app/api/canvas/generation-tasks/route.ts
@ -1,3 +1,4 @@
+import { audioRequestSchema } from "@/lib/video-canvas/audio-contract";
 import { canResumeCanvasTask } from "@/lib/video-canvas/task-recovery";
 import { imageOperationSchema } from "@/lib/video-canvas/image-operation";
 import { after, NextResponse } from "next/server";
@ -52,6 +53,8 @@ export async function POST(request: Request) {
        (parsed.data.attachment.type === "story-prompts" && parsed.data.kind !== "analysis"))
    )
      throw Object.assign(Error("脚本结果类型不匹配"), { status: 400 });
+    if (parsed.data.kind === "audio" && !audioRequestSchema.safeParse(parsed.data.params).success)
+      throw Object.assign(Error("音频参数不合法"), { status: 400 });
    const userId = await owner(parsed.data.projectId);
    const task = await createCanvasGenerationTask(userId, parsed.data);
    kick(userId, parsed.data.projectId);
--- a/components/canvas/AudioGenerationComposer.tsx
+++ b/components/canvas/AudioGenerationComposer.tsx
@ -0,0 +1,306 @@
+"use client";
+import { useEffect, useState } from "react";
+import {
+  audioContract,
+  audioRequestSchema,
+  buildAudioRequest,
+  type AudioRequest,
+} from "@/lib/video-canvas/audio-contract";
+type Model = { id: string; modelId: string; name: string; kind: string };
+export function AudioGenerationComposer({
+  initial,
+  busy,
+  onSubmit,
+  onUpload,
+  onDraft,
+}: {
+  initial?: Record<string, unknown>;
+  busy: boolean;
+  onSubmit: (params: AudioRequest) => Promise<unknown>;
+  onUpload: (file: File) => Promise<void>;
+  onDraft: (params: Record<string, unknown>) => void;
+}) {
+  const [models, setModels] = useState<Model[]>([]),
+    [error, setError] = useState(""),
+    [loading, setLoading] = useState(true),
+    [sending, setSending] = useState(false);
+  const [draft, setDraft] = useState<Record<string, any>>(() => ({
+    mode: "tts",
+    text: "",
+    voice: "Cherry",
+    format: "wav",
+    language: "Chinese",
+    instrumental: true,
+    gender: "female",
+    rate: 1,
+    pitch: 1,
+    volume: 50,
+    ...initial,
+  }));
+  const patch = (value: Record<string, unknown>) => {
+    const next = { ...draft, ...value };
+    setDraft(next);
+    onDraft(next);
+  };
+  useEffect(() => {
+    const controller = new AbortController();
+    fetch("/api/canvas/audio/models", { signal: controller.signal })
+      .then(async (r) => {
+        const data = await r.json();
+        if (!r.ok) throw Error(data.error || "音频模型读取失败");
+        setModels(data.models || []);
+      })
+      .catch((e) => {
+        if (!controller.signal.aborted) setError(e.message);
+      })
+      .finally(() => {
+        if (!controller.signal.aborted) setLoading(false);
+      });
+    return () => controller.abort();
+  }, []);
+  const options = models.filter((m) => m.kind === draft.mode),
+    model = options.find((m) => m.id === draft.modelRef) || options[0],
+    contract = model ? audioContract(model.modelId) : undefined;
+  const voice = contract?.voices.some((v) => v.id === draft.voice)
+    ? draft.voice
+    : contract?.voices[0]?.id;
+  const format = contract?.formats.includes(draft.format) ? draft.format : contract?.formats[0];
+  const input = "mt-1 w-full rounded-lg border border-white/15 bg-[#303030] p-2 text-sm";
+  const run = async () => {
+    setError("");
+    setSending(true);
+    try {
+      if (!model) throw Error("当前模式没有已启用的官方模型");
+      const params = audioRequestSchema.parse({
+        ...draft,
+        modelRef: model.id,
+        voice,
+        format,
+        ...(!contract?.controls ? { rate: 1, pitch: 1, volume: 50 } : {}),
+        ...(!contract?.instructions ? { instructions: "" } : {}),
+      });
+      buildAudioRequest(params, model.modelId);
+      patch(params);
+      await onSubmit(params);
+    } catch (e) {
+      setError(e instanceof Error ? e.message : "提交失败");
+    } finally {
+      setSending(false);
+    }
+  };
+  return (
+    <section
+      className="fixed bottom-8 left-1/2 z-[100] max-h-[75vh] w-[min(620px,94vw)] -translate-x-1/2 overflow-y-auto rounded-2xl border border-white/10 bg-[#222] p-5 text-white shadow-xl"
+      data-canvas-overlay="true"
+      aria-label="阿里云音频生成"
+    >
+      <h3 className="font-semibold">语音与音乐 · 阿里云官方</h3>
+      <fieldset disabled={busy || sending} className="mt-3 space-y-3 disabled:opacity-60">
+        <div className="grid grid-cols-2 gap-3">
+          <label>
+            生成类型
+            <select
+              aria-label="音频生成类型"
+              className={input}
+              value={draft.mode}
+              onChange={(e) => patch({ mode: e.target.value, modelRef: "" })}
+            >
+              <option value="tts">文字配音</option>
+              <option value="music">音乐 / 歌曲</option>
+            </select>
+          </label>
+          <label>
+            模型
+            <select
+              aria-label="音频模型"
+              className={input}
+              value={model?.id || ""}
+              onChange={(e) => patch({ modelRef: e.target.value })}
+            >
+              {!options.length ? (
+                <option value="">{loading ? "读取模型中…" : "暂无已开通模型"}</option>
+              ) : (
+                options.map((m) => (
+                  <option key={m.id} value={m.id}>
+                    {m.name}
+                  </option>
+                ))
+              )}
+            </select>
+          </label>
+        </div>
+        <label className="block">
+          {draft.mode === "tts" ? "配音文本（最多 600 字）" : "音乐描述"}
+          <textarea
+            aria-label="音频内容"
+            className={input}
+            rows={4}
+            maxLength={draft.mode === "tts" ? 600 : 2000}
+            value={draft.text}
+            onChange={(e) => patch({ text: e.target.value })}
+          />
+        </label>
+        {draft.mode === "tts" ? (
+          <>
+            <div className="grid grid-cols-2 gap-3">
+              <label>
+                音色
+                <select
+                  aria-label="配音音色"
+                  className={input}
+                  value={voice || ""}
+                  onChange={(e) => patch({ voice: e.target.value })}
+                >
+                  {contract?.voices.map((v) => (
+                    <option key={v.id} value={v.id}>
+                      {v.name}
+                    </option>
+                  ))}
+                </select>
+              </label>
+              <label>
+                语言
+                <select
+                  aria-label="配音语言"
+                  className={input}
+                  value={draft.language}
+                  onChange={(e) => patch({ language: e.target.value })}
+                >
+                  <option value="Chinese">中文</option>
+                  <option value="English">英文</option>
+                  <option value="Auto">自动 / 中英混合</option>
+                </select>
+              </label>
+            </div>
+            {contract?.instructions ? (
+              <label className="block">
+                声音风格指令
+                <input
+                  className={input}
+                  aria-label="声音风格指令"
+                  placeholder="例如：温暖自然，语速略慢"
+                  maxLength={1200}
+                  value={draft.instructions || ""}
+                  onChange={(e) => patch({ instructions: e.target.value })}
+                />
+              </label>
+            ) : null}
+            {contract?.controls ? (
+              <div className="grid grid-cols-3 gap-2">
+                {(["rate", "pitch", "volume"] as const).map((key, i) => (
+                  <label key={key}>
+                    {["语速", "音调", "音量"][i]}
+                    <input
+                      aria-label={["语速", "音调", "音量"][i]}
+                      className={input}
+                      type="number"
+                      min={key === "volume" ? 0 : 0.5}
+                      max={key === "volume" ? 100 : 2}
+                      step={key === "volume" ? 1 : 0.1}
+                      value={draft[key]}
+                      onChange={(e) => patch({ [key]: Number(e.target.value) })}
+                    />
+                  </label>
+                ))}
+              </div>
+            ) : null}
+          </>
+        ) : (
+          <>
+            <label className="flex gap-2">
+              <input
+                type="checkbox"
+                checked={draft.instrumental}
+                onChange={(e) => patch({ instrumental: e.target.checked })}
+              />
+              纯音乐（无人声）
+            </label>
+            {!draft.instrumental ? (
+              <>
+                <label className="block">
+                  自定义歌词（选填，填写后优先于描述）
+                  <textarea
+                    aria-label="音乐歌词"
+                    className={input}
+                    rows={3}
+                    maxLength={2000}
+                    value={draft.lyrics || ""}
+                    onChange={(e) => patch({ lyrics: e.target.value })}
+                  />
+                </label>
+                <label>
+                  演唱声音
+                  <select
+                    className={input}
+                    aria-label="演唱声音"
+                    value={draft.gender}
+                    onChange={(e) => patch({ gender: e.target.value })}
+                  >
+                    <option value="female">女声</option>
+                    <option value="male">男声</option>
+                  </select>
+                </label>
+              </>
+            ) : null}
+            {!options.length ? (
+              <p className="text-sm text-amber-200">
+                Fun-Music 需要阿里云邀测权限，开通并启用模型后即可生成。
+              </p>
+            ) : null}
+          </>
+        )}
+        <div className="flex items-end gap-3">
+          <label>
+            格式
+            <select
+              aria-label="音频格式"
+              className={input}
+              value={format || "wav"}
+              onChange={(e) => patch({ format: e.target.value })}
+            >
+              {(contract?.formats || ["wav"]).map((f) => (
+                <option key={f} value={f}>
+                  {f.toUpperCase()}
+                </option>
+              ))}
+            </select>
+          </label>
+          <button
+            type="button"
+            disabled={!model || loading}
+            className="rounded-lg bg-white px-4 py-2 text-black disabled:opacity-40"
+            onClick={() => void run()}
+          >
+            {busy || sending ? "处理中…" : "生成音频"}
+          </button>
+          <label className="cursor-pointer rounded-lg border border-white/20 px-3 py-2">
+            上传音频
+            <input
+              type="file"
+              accept="audio/*"
+              className="hidden"
+              onChange={async (e) => {
+                const file = e.target.files?.[0];
+                e.target.value = "";
+                if (!file) return;
+                setSending(true);
+                try {
+                  await onUpload(file);
+                } catch (e) {
+                  setError(e instanceof Error ? e.message : "上传失败");
+                } finally {
+                  setSending(false);
+                }
+              }}
+            />
+          </label>
+        </div>
+      </fieldset>
+      {error ? (
+        <p role="alert" className="mt-3 text-sm text-red-300">
+          {error}
+        </p>
+      ) : null}
+    </section>
+  );
+}
--- a/components/canvas/LiblibVideoCanvasReplica.tsx
+++ b/components/canvas/LiblibVideoCanvasReplica.tsx
@ -1,4 +1,5 @@
 "use client";
+import { AudioGenerationComposer } from "./AudioGenerationComposer";
 import { VideoCanvasConflictNotice } from "./VideoCanvasConflictNotice";
 import { normalizeVideoAnalysisRows } from "@/lib/video-canvas/video-story";
 import dynamic from "next/dynamic";
@ -8722,6 +8723,13 @@ export function LiblibVideoCanvasReplica({
      setNodeContextMenu(null);
      setConnectionMenu(null);

+      if (node.inputType === "audio" && node.draftValues?.audioParams) {
+        await generationTasks.submit(
+          node, "audio", node.draftValues.audioParams as Record<string, unknown>,
+          edges.filter(edge => edge.target === nodeId),
+        );
+        return;
+      }
      if (getDraftString(node, "action") === "VIDEO_STORY_ANALYSIS") {
        await requestVideoAnalysis(getDraftString(node, "sourceVideoNodeId"), true);
        return;
@ -10524,7 +10532,17 @@ export function LiblibVideoCanvasReplica({
      ) : null}

      {selectedNode && selectedTargetModuleKind === "audio" && selectedTargetAnchor ? (
-        <AudioTargetModuleComposer anchor={selectedTargetAnchor} onUpload={async file => {
+        <AudioGenerationComposer
+          key={selectedNode.id}
+          initial={selectedNode.draftValues?.audioParams as Record<string, unknown> | undefined}
+          busy={selectedNode.draftValues?.generationStatus === "pending"}
+          onDraft={params => setNodes(current => current.map(node => node.id === selectedNode.id
+            ? { ...node, draftValues: { ...node.draftValues, audioParams: params } } : node))}
+          onSubmit={params => generationTasks.submit({
+            ...selectedNode, kind: "media", inputType: "audio",
+            draftValues: { ...selectedNode.draftValues, audioParams: params },
+          }, "audio", params, edges.filter(edge => edge.target === selectedNode.id))}
+          onUpload={async file => {
          const form = new FormData(); form.append("file", file);
          const response = await fetch("/api/canvas/media/upload", {method:"POST",body:form,signal:AbortSignal.timeout(120000)});
          const result = await response.json(); const url = result.url || result.data?.url;
@ -10944,6 +10962,19 @@ export function LiblibVideoCanvasReplica({
      {directorConsoleNode ? (
        <DirectorStudioDialog
          initialScene={directorConsoleNode.draftValues?.directorScene}
+          reconstructionTask={generationTasks.tasks.filter(task =>
+            task.request.params.sourceDirectorNodeId === directorConsoleNode.id &&
+            task.request.params.mode === "director-reconstruct").slice(-1)[0]}
+          onReconstruct={async ({ imageUrl, prompt }) => {
+            const node = createTargetModuleNode("director", directorConsoleNode.x + 600,
+              directorConsoleNode.y, nodes.length + 1);
+            node.title = "AI 识图重建布景";
+            await generationTasks.submit(node, "analysis", {
+              mode: "director-reconstruct", prompt, imageUrls: [imageUrl],
+              preferredModelRef: videoCanvasFunctionConfig.defaultMultimodalModel,
+              sourceDirectorNodeId: directorConsoleNode.id,
+            }, [{ id: makeId("edge"), source: directorConsoleNode.id, target: node.id }]);
+          }}
          onChange={scene => setNodes(current => current.map(node => node.id === directorConsoleNode.id ? {...node, draftValues: {...node.draftValues, directorScene: scene}} : node))}
          onSend={snapshots => sendDirectorSnapshotsToCanvas(directorConsoleNode, snapshots)}
          onSendMotion={(snapshot, motion) => {
@ -13505,49 +13536,6 @@ function TextTargetModuleComposer({
  );
 }

-function AudioTargetModuleComposer({
-  anchor,
-  onUpload,
-}: {
-  anchor: PresetControlAnchor;
-  onUpload: (file: File) => Promise<void>;
-}) {
-  const [busy, setBusy] = useState(false);
-  return (
-    <div
-      className="fixed bottom-8 left-1/2 z-[100] w-[min(520px,90vw)] -translate-x-1/2 rounded-2xl border border-white/10 bg-[#222] p-5 text-white shadow-xl"
-      data-canvas-overlay="true"
-    >
-      <h3 className="text-sm font-semibold">音频素材</h3>
-      <p className="my-3 text-sm leading-6 text-white/60">
-        当前尚未接通语音与音乐生成线路。可上传音频用于预览，或从视频工具提取音轨，再连接到支持音频参考的视频节点。
-      </p>
-      <label className="inline-flex cursor-pointer rounded-lg bg-white px-4 py-2 text-sm text-black">
-        {busy ? "上传中…" : "上传音频"}
-        <input
-          type="file"
-          accept="audio/*"
-          className="hidden"
-          disabled={busy}
-          onChange={async (e) => {
-            const file = e.target.files?.[0];
-            e.target.value = "";
-            if (!file) return;
-            setBusy(true);
-            try {
-              await onUpload(file);
-            } catch (e) {
-              toast.error(e instanceof Error ? e.message : "上传失败");
-            } finally {
-              setBusy(false);
-            }
-          }}
-        />
-      </label>
-    </div>
-  );
-}
-
 function ReferenceSelectionToolbar({
  onExit,
  onReturn,
--- a/components/canvas/director/DirectorMotionPanel.tsx
+++ b/components/canvas/director/DirectorMotionPanel.tsx
@ -13,6 +13,8 @@ import {
 } from "@/lib/video-canvas/director-motion";
 export function DirectorMotionPanel({
  motion,
+  hasObjectMotion = false,
+  externalSeek,
  compact = false,
  pose,
  disabled,
@ -26,6 +28,8 @@ export function DirectorMotionPanel({
  onOverview,
 }: {
  compact?: boolean;
+  hasObjectMotion?: boolean;
+  externalSeek?: { time: number; editing?: boolean };
  motion: DirectorMotion;
  pose: DirectorPose;
  disabled: boolean;
@ -136,6 +140,13 @@ export function DirectorMotionPanel({
    previousPose.current = signature;
    if (autoKeyframe && changed && !preview && !disabled) add();
  }, [pose, autoKeyframe, preview, disabled]);
+  useEffect(() => {
+    if (!externalSeek) return;
+    setPlaying(false);
+    setTime(externalSeek.time);
+    timeRef.current = externalSeek.time;
+    setPreview(!externalSeek.editing);
+  }, [externalSeek]);
  const downloadJson = () => {
    const url = URL.createObjectURL(
      new Blob([JSON.stringify({ version: 1, motion }, null, 2)], {
@ -222,7 +233,7 @@ export function DirectorMotionPanel({
        <button
          type="button"
          className={button}
-          disabled={motion.keyframes.length < 2}
+          disabled={motion.keyframes.length < 2 && !hasObjectMotion}
          onClick={() => {
            edit();
            onOverview();
@ -253,7 +264,7 @@ export function DirectorMotionPanel({
          <button
            type="button"
            className={button}
-            disabled={motion.keyframes.length < 2}
+            disabled={motion.keyframes.length < 2 && !hasObjectMotion}
            onClick={() => {
              if (playing) {
                setPlaying(false);
@ -397,7 +408,7 @@ export function DirectorMotionPanel({
        <button
          type="button"
          className={`${button} w-full`}
-          disabled={motion.keyframes.length < 2}
+          disabled={motion.keyframes.length < 2 && !hasObjectMotion}
          onClick={() => {
            edit();
            onExport();
--- a/components/canvas/director/DirectorObjectTimeline.tsx
+++ b/components/canvas/director/DirectorObjectTimeline.tsx
@ -0,0 +1,256 @@
+"use client";
+import { useState } from "react";
+import { toast } from "sonner";
+import type { DirectorScene } from "@/lib/video-canvas/director-scene";
+import { recordObjectFrames } from "@/lib/video-canvas/director-object-tracks";
+export function DirectorObjectTimeline({
+  scene,
+  time,
+  selected,
+  disabled,
+  preview,
+  auto,
+  onAuto,
+  onChange,
+  onSeek,
+  onEdit,
+  onSelect,
+}: {
+  scene: DirectorScene;
+  time: number;
+  selected: string | null;
+  disabled: boolean;
+  preview: boolean;
+  auto: boolean;
+  onAuto: (v: boolean) => void;
+  onChange: (s: DirectorScene) => void;
+  onSeek: (t: number) => void;
+  onEdit: () => void;
+  onSelect: (id: string) => void;
+}) {
+  const [unit, setUnit] = useState("seconds"),
+    [zoom, setZoom] = useState(1);
+  const factor = unit === "frames" ? 30 : 1,
+    track = selected ? scene.objectTracks?.[selected] : undefined;
+  const button = "rounded border border-white/20 px-2 py-1 text-xs disabled:opacity-40",
+    field = "rounded bg-[#303030] p-1 text-xs";
+  return (
+    <fieldset
+      disabled={disabled}
+      className="space-y-3 border-t border-white/15 p-3"
+      aria-label="对象多轨时间轴"
+    >
+      <legend className="px-2 text-sm">对象多轨 · 与相机同步</legend>
+      <div className="flex flex-wrap items-center gap-3">
+        <label>
+          单位{" "}
+          <select
+            aria-label="时间轴单位"
+            value={unit}
+            className={field}
+            onChange={(e) => setUnit(e.target.value)}
+          >
+            <option value="seconds">秒</option>
+            <option value="frames">帧（30 fps）</option>
+          </select>
+        </label>
+        <label>
+          缩放{" "}
+          <select
+            aria-label="对象时间轴缩放"
+            value={zoom}
+            className={field}
+            onChange={(e) => setZoom(Number(e.target.value))}
+          >
+            {[1, 2, 4, 8].map((n) => (
+              <option key={n} value={n}>
+                {n}×
+              </option>
+            ))}
+          </select>
+        </label>
+        <span>
+          {(time * factor).toFixed(unit === "frames" ? 0 : 2)} /{" "}
+          {(scene.motion.duration * factor).toFixed(unit === "frames" ? 0 : 2)}{" "}
+          {unit === "frames" ? "帧" : "秒"}
+        </span>
+        <label>
+          <input type="checkbox" checked={auto} onChange={(e) => onAuto(e.target.checked)} />{" "}
+          自动记录对象帧
+        </label>
+        <button type="button" className={button} onClick={onEdit}>
+          编辑当前时间对象
+        </button>
+        <button
+          type="button"
+          className={button}
+          disabled={preview || !selected}
+          onClick={() => {
+            try {
+              onChange(recordObjectFrames(scene, [selected!], time, () => crypto.randomUUID()));
+            } catch (e) {
+              toast.error((e as Error).message);
+            }
+          }}
+        >
+          记录所选对象帧
+        </button>
+      </div>
+      <div className="overflow-x-auto">
+        <div style={{ minWidth: `${zoom * 100}%` }}>
+          <input
+            aria-label="对象播放头"
+            type="range"
+            min={0}
+            max={scene.motion.duration}
+            step={1 / 30}
+            value={time}
+            className="w-full"
+            onChange={(e) => onSeek(Number(e.target.value))}
+          />
+          {scene.objects.map((o) => {
+            const t = scene.objectTracks?.[o.id];
+            return (
+              <div
+                key={o.id}
+                className="grid grid-cols-[150px_60px_1fr] items-center gap-2 border-t border-white/10 py-2 text-xs"
+              >
+                <button
+                  type="button"
+                  className={`truncate text-left ${selected === o.id ? "text-amber-300" : ""}`}
+                  onClick={() => onSelect(o.id)}
+                >
+                  {o.name}
+                </button>
+                <label>
+                  <input
+                    type="checkbox"
+                    aria-label={`${o.name}启用轨道`}
+                    checked={!t?.muted}
+                    onChange={(e) =>
+                      onChange({
+                        ...scene,
+                        objectTracks: {
+                          ...scene.objectTracks,
+                          [o.id]: {
+                            keyframes: t?.keyframes || [],
+                            easing: t?.easing || "linear",
+                            muted: !e.target.checked,
+                          },
+                        },
+                      })
+                    }
+                  />
+                  启用
+                </label>
+                <div className="relative h-6 rounded bg-white/5">
+                  {t?.keyframes.map((k, i) => (
+                    <button
+                      key={k.id}
+                      type="button"
+                      aria-label={`${o.name}关键帧${i + 1}`}
+                      className="absolute top-0 -translate-x-1/2 text-lg text-amber-300"
+                      style={{ left: `${(k.time / scene.motion.duration) * 100}%` }}
+                      onClick={() => {
+                        onSelect(o.id);
+                        onSeek(k.time);
+                      }}
+                    >
+                      ◆
+                    </button>
+                  ))}
+                </div>
+              </div>
+            );
+          })}
+        </div>
+      </div>
+      {track && selected ? (
+        <div className="space-y-2">
+          <label>
+            对象插值{" "}
+            <select
+              aria-label="对象插值"
+              value={track.easing}
+              className={field}
+              onChange={(e) =>
+                onChange({
+                  ...scene,
+                  objectTracks: {
+                    ...scene.objectTracks,
+                    [selected]: { ...track, easing: e.target.value as "smooth" | "linear" },
+                  },
+                })
+              }
+            >
+              <option value="linear">匀速</option>
+              <option value="smooth">缓入缓出</option>
+            </select>
+          </label>
+          <div className="flex flex-wrap gap-2">
+            {track.keyframes.map((k, i) => (
+              <label className="flex items-center gap-1" key={k.id}>
+                帧 {i + 1}
+                <input
+                  className={`${field} w-20`}
+                  aria-label={`对象关键帧${i + 1}时间`}
+                  type="number"
+                  min={0}
+                  max={scene.motion.duration * factor}
+                  step={unit === "frames" ? 1 : 0.01}
+                  value={Number((k.time * factor).toFixed(2))}
+                  onChange={(e) => {
+                    const next = Number(e.target.value) / factor;
+                    if (
+                      !Number.isFinite(next) ||
+                      next < 0 ||
+                      next > scene.motion.duration ||
+                      track.keyframes.some(
+                        (other) => other.id !== k.id && Math.abs(other.time - next) < 0.001,
+                      )
+                    )
+                      return;
+                    onChange({
+                      ...scene,
+                      objectTracks: {
+                        ...scene.objectTracks,
+                        [selected]: {
+                          ...track,
+                          keyframes: track.keyframes
+                            .map((old) => (old.id === k.id ? { ...old, time: next } : old))
+                            .sort((a, b) => a.time - b.time),
+                        },
+                      },
+                    });
+                  }}
+                />
+                <button
+                  type="button"
+                  className={button}
+                  aria-label={`删除对象关键帧${i + 1}`}
+                  onClick={() =>
+                    onChange({
+                      ...scene,
+                      objectTracks: {
+                        ...scene.objectTracks,
+                        [selected]: {
+                          ...track,
+                          keyframes: track.keyframes.filter((old) => old.id !== k.id),
+                        },
+                      },
+                    })
+                  }
+                >
+                  删除
+                </button>
+              </label>
+            ))}
+          </div>
+        </div>
+      ) : null}
+      <p className="text-xs text-white/50">
+        拖动播放头预览全部对象；点击“编辑当前时间对象”后调整场景对象，再记录关键帧。真实视频导出使用相同轨道。
+      </p>
+    </fieldset>
+  );
+}
--- a/components/canvas/director/DirectorReconstructionPanel.tsx
+++ b/components/canvas/director/DirectorReconstructionPanel.tsx
@ -0,0 +1,118 @@
+"use client";
+import { useState } from "react";
+import { toast } from "sonner";
+import { readDirectorScene, type DirectorScene } from "@/lib/video-canvas/director-scene";
+import { insertReconstructedScene } from "@/lib/video-canvas/director-reconstruction";
+export type ReconstructionTask = {
+  status: string;
+  error?: string | null;
+  result?: Record<string, unknown> | null;
+};
+export function DirectorReconstructionPanel({
+  scene,
+  onChange,
+  onSubmit,
+  task,
+  disabled,
+}: {
+  scene: DirectorScene;
+  onChange: (s: DirectorScene) => void;
+  onSubmit: (input: { imageUrl: string; prompt: string }) => Promise<unknown>;
+  task?: ReconstructionTask;
+  disabled: boolean;
+}) {
+  const [file, setFile] = useState<File | null>(null),
+    [prompt, setPrompt] = useState(""),
+    [busy, setBusy] = useState(false);
+  const waiting = task && ["pending", "running"].includes(task.status);
+  const result = task?.status === "succeeded" ? task.result : undefined;
+  return (
+    <fieldset
+      disabled={disabled || busy}
+      className="space-y-2 rounded border border-white/15 p-3 text-xs"
+    >
+      <legend>AI 识图重建布景</legend>
+      <p className="text-white/60">
+        从参考图识别人物、物体与机位，生成可编辑的简化几何布景；不包含精细网格和纹理。
+      </p>
+      <input
+        aria-label="重建参考图"
+        type="file"
+        accept="image/png,image/jpeg,image/webp"
+        onChange={(e) => setFile(e.target.files?.[0] || null)}
+      />
+      <textarea
+        aria-label="重建要求"
+        className="w-full rounded bg-[#303030] p-2"
+        value={prompt}
+        onChange={(e) => setPrompt(e.target.value)}
+        placeholder="需要保留的主体、构图与大致尺寸"
+        maxLength={2000}
+      />
+      <button
+        type="button"
+        disabled={!file || !!waiting}
+        className="rounded bg-amber-400 px-3 py-2 text-black disabled:opacity-40"
+        onClick={async () => {
+          if (!file) return;
+          if (file.size > 20 * 1024 * 1024) {
+            toast.error("参考图需小于 20 MiB");
+            return;
+          }
+          setBusy(true);
+          try {
+            const form = new FormData();
+            form.append("file", file);
+            form.append("purpose", "director-studio");
+            const response = await fetch("/api/files/upload", {
+              method: "POST",
+              body: form,
+              signal: AbortSignal.timeout(120000),
+            });
+            const data = await response.json();
+            const url = data.url || data.data?.url;
+            if (!response.ok || !url) throw Error(data.error || "参考图上传失败");
+            await onSubmit({ imageUrl: url, prompt });
+            toast.success("重建任务已提交，结果将保存为独立导演台节点");
+          } catch (e) {
+            toast.error(e instanceof Error ? e.message : "重建失败");
+          } finally {
+            setBusy(false);
+          }
+        }}
+      >
+        {busy ? "正在提交…" : waiting ? "正在重建…" : "识图重建"}
+      </button>
+      {task?.error ? <p role="alert">{task.error}</p> : null}
+      {result?.directorScene ? (
+        <div>
+          <p className="my-2">{String(result.summary || "布景已生成")}</p>
+          {[false, true].map((replace) => (
+            <button
+              type="button"
+              key={String(replace)}
+              className="mr-2 rounded border border-white/20 px-2 py-1"
+              onClick={() => {
+                try {
+                  onChange(
+                    insertReconstructedScene(
+                      scene,
+                      readDirectorScene(result.directorScene),
+                      replace,
+                      () => crypto.randomUUID(),
+                    ),
+                  );
+                  toast.success(replace ? "已替换场景对象" : "已插入场景对象");
+                } catch (e) {
+                  toast.error((e as Error).message);
+                }
+              }}
+            >
+              {replace ? "替换场景对象" : "插入当前场景"}
+            </button>
+          ))}
+        </div>
+      ) : null}
+    </fieldset>
+  );
+}
--- a/components/canvas/director/DirectorStudioDialog.tsx
+++ b/components/canvas/director/DirectorStudioDialog.tsx
@ -21,6 +21,15 @@ import {
  type DirectorObject,
  type DirectorVector,
 } from "@/lib/video-canvas/director-scene";
+import { DirectorObjectTimeline } from "./DirectorObjectTimeline";
+import {
+  DirectorReconstructionPanel,
+  type ReconstructionTask,
+} from "./DirectorReconstructionPanel";
+import {
+  recordObjectFrames,
+  sampleDirectorObjects,
+} from "@/lib/video-canvas/director-object-tracks";
 import { DirectorMotionPanel } from "./DirectorMotionPanel";
 import {
  sampleDirectorMotion,
@ -54,7 +63,11 @@ export function DirectorStudioDialog({
  onSend,
  onSendMotion,
  onSendVideo,
+  onReconstruct,
+  reconstructionTask,
 }: {
+  onReconstruct?: (input: { imageUrl: string; prompt: string }) => Promise<unknown>;
+  reconstructionTask?: ReconstructionTask;
  initialScene: unknown;
  onChange: (scene: DirectorScene) => void;
  onClose: () => void;
@ -78,7 +91,16 @@ export function DirectorStudioDialog({
  const [cameraView, setCameraView] = useState(false);
  const [busy, setBusy] = useState(false);
  const viewport = useRef<DirectorViewportHandle>(null);
-  const [previewTime, setPreviewTime] = useState<number | null>(null);
+  const timelineDrawer = useRef<HTMLDetailsElement>(null);
+  const reconstructionAnchor = useRef<HTMLDivElement>(null);
+  const [previewTime, setPreviewState] = useState<number | null>(null);
+  const [timelineTime, setTimelineTime] = useState(0);
+  const [externalSeek, setExternalSeek] = useState<{ time: number; editing?: boolean }>();
+  const [autoObjectFrame, setAutoObjectFrame] = useState(false);
+  const setPreviewTime = (time: number | null) => {
+    setPreviewState(time);
+    if (time !== null) setTimelineTime(time);
+  };
  const [exportProgress, setExportProgress] = useState<number | null>(null);
  const recording = useRef<AbortController | null>(null);
  useEffect(() => () => recording.current?.abort(), []);
@ -135,7 +157,7 @@ export function DirectorStudioDialog({
      };
      const pose = sampleDirectorMotion(motion, 0);
      if (!pose || motion.keyframes.length < 2) throw Error("至少需要两个关键帧");
-      const image = await viewport.current!.capture(pose);
+      const image = await viewport.current!.capture(pose, 0);
      const url = await upload(
        new File([image.blob], `director-motion-${Date.now()}.png`, {
          type: "image/png",
@ -166,11 +188,23 @@ export function DirectorStudioDialog({
  const camera = resolveDirectorCamera(scene);
  const angles = cameraAngles(camera);
  const object = scene.objects.find((o) => o.id === selected);
-  const patchObject = (id: string, patch: Partial<DirectorObject>) =>
-    commit({
+  const patchObject = (id: string, patch: Partial<DirectorObject>) => {
+    let next = {
      ...current.current,
      objects: current.current.objects.map((o) => (o.id === id ? { ...o, ...patch } : o)),
-    });
+    };
+    try {
+      if (
+        autoObjectFrame &&
+        previewTime === null &&
+        ["position", "rotation", "scale", "visible"].some((key) => Object.hasOwn(patch, key))
+      )
+        next = recordObjectFrames(next, [id], timelineTime, () => crypto.randomUUID());
+      commit(next);
+    } catch (error) {
+      toast.error((error as Error).message);
+    }
+  };
  const patchCamera = (patch: Partial<typeof camera>) =>
    commit({
      ...current.current,
@ -338,8 +372,18 @@ export function DirectorStudioDialog({
                ))}
              </select>
            </label>
-            <button className={button} disabled title="尚未接入图片识别与场景重建">
-              AI 识图导入（未接通）
+            <button
+              className={button}
+              disabled={!onReconstruct || busy}
+              onClick={() => {
+                if (timelineDrawer.current) timelineDrawer.current.open = true;
+                reconstructionAnchor.current?.scrollIntoView({
+                  block: "nearest",
+                  behavior: "smooth",
+                });
+              }}
+            >
+              AI 识图重建布景
            </button>
          </div>
          <p className="mt-4 text-xs leading-5 text-white/45">
@ -789,6 +833,7 @@ export function DirectorStudioDialog({
        </aside>
      </div>
      <details
+        ref={timelineDrawer}
        open
        className="max-h-[32vh] shrink-0 overflow-auto border-t border-white/15 bg-[#1c1c1c] px-4 pb-3"
      >
@ -798,6 +843,10 @@ export function DirectorStudioDialog({
        <DirectorMotionPanel
          key={scene.activeCameraId}
          compact
+          hasObjectMotion={Object.values(scene.objectTracks || {}).some(
+            (t) => !t.muted && t.keyframes.length >= 2,
+          )}
+          externalSeek={externalSeek}
          motion={scene.motion}
          pose={camera}
          disabled={busy}
@ -813,6 +862,40 @@ export function DirectorStudioDialog({
            viewport.current?.showPath();
          }}
        />
+        <DirectorObjectTimeline
+          scene={scene}
+          time={timelineTime}
+          selected={selected}
+          disabled={busy}
+          preview={previewTime !== null}
+          auto={autoObjectFrame}
+          onAuto={setAutoObjectFrame}
+          onChange={commit}
+          onSelect={setSelected}
+          onSeek={(time) => {
+            setExternalSeek({ time });
+            setPreviewTime(time);
+          }}
+          onEdit={() => {
+            commit({
+              ...current.current,
+              objects: sampleDirectorObjects(current.current, timelineTime),
+            });
+            setPreviewTime(null);
+            setExternalSeek({ time: timelineTime, editing: true });
+          }}
+        />
+        {onReconstruct ? (
+          <div ref={reconstructionAnchor}>
+            <DirectorReconstructionPanel
+              scene={scene}
+              onChange={commit}
+              onSubmit={onReconstruct}
+              task={reconstructionTask}
+              disabled={busy}
+            />
+          </div>
+        ) : null}
      </details>
      <input
        ref={backgroundInput}
--- a/components/canvas/director/DirectorViewport.tsx
+++ b/components/canvas/director/DirectorViewport.tsx
@ -1,4 +1,5 @@
 "use client";
+import { sampleDirectorObjects } from "@/lib/video-canvas/director-object-tracks";
 import { resolveDirectorCamera, directorBodies } from "@/lib/video-canvas/director-extensions";
 import { directorPoses } from "@/lib/video-canvas/director-poses";
 import { forwardRef, useEffect, useImperativeHandle, useRef, useState } from "react";
@ -17,7 +18,10 @@ import { recordDirectorVideo } from "@/lib/video-canvas/director-recording";
 export type DirectorViewportHandle = {
  showPath: () => void;
  record: (signal: AbortSignal, onProgress: (time: number) => void) => Promise<Blob>;
-  capture: (pose?: DirectorPose) => Promise<{ blob: Blob; width: number; height: number }>;
+  capture: (
+    pose?: DirectorPose,
+    time?: number,
+  ) => Promise<{ blob: Blob; width: number; height: number }>;
 };
 type Props = {
  scene: DirectorScene;
@ -65,6 +69,21 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
    const [loading, setLoading] = useState(false);
    const [backgroundPending, setBackgroundPending] = useState(false);
    const [backgroundError, setBackgroundError] = useState("");
+    const applyObjectTime = (time: number | null) => {
+      const objects = sampleDirectorObjects(latest.current.scene, time),
+        r = runtime.current;
+      if (r)
+        for (const item of objects) {
+          const mesh = r.objects.children.find((o) => o.name === item.id);
+          if (mesh) {
+            mesh.position.fromArray(item.position);
+            mesh.rotation.set(...item.rotation);
+            mesh.scale.fromArray(item.scale);
+            mesh.visible = item.visible !== false;
+          }
+        }
+      return { ...latest.current.scene, objects };
+    };
    useEffect(() => {
      if (!host.current) return;
      setError("");
@ -210,7 +229,14 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
          return;
        }
        object.name = item.id;
-        object.visible = item.visible !== false;
+        // A GLB may finish loading after playback has paused at a nonzero time.
+        const sampled =
+          latest.current.previewTime === null
+            ? item
+            : sampleDirectorObjects(latest.current.scene, latest.current.previewTime).find(
+                (o) => o.id === item.id,
+              ) || item;
+        object.visible = sampled.visible !== false;
        object.userData.directorSignature = JSON.stringify([
          item.kind,
          item.url,
@ -218,9 +244,9 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
          item.color,
          item.body,
        ]);
-        object.position.fromArray(item.position);
-        object.rotation.set(...item.rotation);
-        object.scale.fromArray(item.scale);
+        object.position.fromArray(sampled.position);
+        object.rotation.set(...sampled.rotation);
+        object.scale.fromArray(sampled.scale);
        r.objects.add(object);
        if (
          item.id === latest.current.selected &&
@ -477,35 +503,64 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
      dispose(r.path);
      r.path.clear();
      const motion = props.scene.motion;
-      if (motion.keyframes.length < 2) return;
-      const points = Array.from(
-        { length: 121 },
-        (_, i) =>
-          new THREE.Vector3(...sampleDirectorMotion(motion, (motion.duration * i) / 120)!.position),
-      );
-      r.path.add(
-        new THREE.Line(
-          new THREE.BufferGeometry().setFromPoints(points),
-          new THREE.LineBasicMaterial({ color: 0xfbbf24 }),
-        ),
-      );
-      for (const k of motion.keyframes) {
-        const dot = new THREE.Mesh(
-          new THREE.SphereGeometry(0.08, 10, 8),
-          new THREE.MeshBasicMaterial({ color: 0xfbbf24 }),
+      const tracks = [
+        ...(motion.keyframes.length >= 2
+          ? [
+              {
+                color: 0xfbbf24,
+                keys: motion.keyframes,
+                points: Array.from(
+                  { length: 121 },
+                  (_, i) =>
+                    new THREE.Vector3(
+                      ...sampleDirectorMotion(motion, (motion.duration * i) / 120)!.position,
+                    ),
+                ),
+              },
+            ]
+          : []),
+        ...Object.entries(props.scene.objectTracks || {})
+          .filter(([, t]) => !t.muted && t.keyframes.length >= 2)
+          .map(([id, t]) => ({
+            color: 0x67e8f9,
+            keys: t.keyframes,
+            points: Array.from(
+              { length: 121 },
+              (_, i) =>
+                new THREE.Vector3(
+                  ...sampleDirectorObjects(props.scene, (motion.duration * i) / 120).find(
+                    (o) => o.id === id,
+                  )!.position,
+                ),
+            ),
+          })),
+      ];
+      for (const track of tracks) {
+        r.path.add(
+          new THREE.Line(
+            new THREE.BufferGeometry().setFromPoints(track.points),
+            new THREE.LineBasicMaterial({ color: track.color }),
+          ),
        );
-        dot.position.fromArray(k.position);
-        r.path.add(dot);
+        for (const k of track.keys) {
+          const dot = new THREE.Mesh(
+            new THREE.SphereGeometry(0.08, 10, 8),
+            new THREE.MeshBasicMaterial({ color: track.color }),
+          );
+          dot.position.fromArray(k.position);
+          r.path.add(dot);
+        }
      }
-    }, [props.scene.motion, ready]);
+    }, [props.scene.motion, props.scene.objectTracks, props.scene.objects, ready]);
    useEffect(() => {
      const r = runtime.current;
      if (!r) return;
      r.path.visible = !props.locked && !props.cameraView && props.previewTime === null;
-      if (props.locked || props.previewTime === null) return;
+      if (props.locked) return;
+      const sampledScene = applyObjectTime(props.previewTime);
+      if (props.previewTime === null) return;
      const sampled = sampleDirectorMotion(props.scene.motion, props.previewTime);
-      if (!sampled) return;
-      const pose = resolveDirectorCamera(props.scene, sampled);
+      const pose = resolveDirectorCamera(sampledScene, sampled || undefined);
      r.camera.position.fromArray(pose.position);
      r.orbit.target.fromArray(pose.target);
      r.camera.fov = pose.fov;
@ -517,6 +572,7 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
      props.scene.motion,
      props.scene.cameras,
      props.scene.objects,
+      props.scene.objectTracks,
      props.cameraView,
      props.locked,
      ready,
@ -533,6 +589,9 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
          const points = latest.current.scene.motion.keyframes.map(
            (k) => new THREE.Vector3(...k.position),
          );
+          for (const track of Object.values(latest.current.scene.objectTracks || {}))
+            if (!track.muted)
+              for (const key of track.keyframes) points.push(new THREE.Vector3(...key.position));
          if (!points.length) return;
          points.push(new THREE.Vector3(0, 1, 0));
          const box = new THREE.Box3().setFromPoints(points),
@ -548,7 +607,13 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
            throw Error(backgroundError || "全景背景仍在加载，请稍后导出");
          const r = runtime.current;
          if (!r || !ready || r.pending || error) throw Error(error || "3D 场景尚未准备好");
-          if (latest.current.scene.motion.keyframes.length < 2) throw Error("至少需要两个关键帧");
+          if (
+            latest.current.scene.motion.keyframes.length < 2 &&
+            !Object.values(latest.current.scene.objectTracks || {}).some(
+              (t) => !t.muted && t.keyframes.length >= 2,
+            )
+          )
+            throw Error("相机或对象轨道至少需要两个关键帧");
          const motion = latest.current.scene.motion;
          const size = r.renderer.getSize(new THREE.Vector2()),
            pixelRatio = r.renderer.getPixelRatio();
@ -576,8 +641,8 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
                if (r.renderer.getContext().isContextLost())
                  throw Error("3D 显示连接中断，已取消视频导出");
                const pose = resolveDirectorCamera(
-                  latest.current.scene,
-                  sampleDirectorMotion(motion, time)!,
+                  applyObjectTime(time),
+                  sampleDirectorMotion(motion, time) || undefined,
                );
                r.camera.position.fromArray(pose.position);
                r.orbit.target.fromArray(pose.target);
@ -589,6 +654,7 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
            });
          } finally {
            if (runtime.current === r) {
+              applyObjectTime(latest.current.previewTime);
              r.renderer.setPixelRatio(pixelRatio);
              r.renderer.setSize(size.x, size.y);
              r.camera.position.copy(position);
@ -602,7 +668,7 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
            }
          }
        },
-        capture: async (pose) => {
+        capture: async (pose, time) => {
          if (backgroundPending || backgroundError)
            throw Error(backgroundError || "全景背景仍在加载，请稍后截图");
          const r = runtime.current;
@ -617,8 +683,9 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
          const visibility = [r.helper.visible, r.grid.visible, r.path.visible];
          const enabled = r.orbit.enabled;
          r.orbit.enabled = false;
+          const capturedScene = applyObjectTime(time ?? latest.current.previewTime);
          if (pose) {
-            pose = resolveDirectorCamera(latest.current.scene, pose);
+            pose = resolveDirectorCamera(capturedScene, pose);
            r.camera.position.fromArray(pose.position);
            r.orbit.target.fromArray(pose.target);
            r.camera.fov = pose.fov;
@ -641,6 +708,7 @@ export const DirectorViewport = forwardRef<DirectorViewportHandle, Props>(
            );
            return { blob, width, height };
          } finally {
+            applyObjectTime(latest.current.previewTime);
            r.renderer.setPixelRatio(pixelRatio);
            r.renderer.setSize(previousSize.x, previousSize.y);
            r.camera.aspect = aspect;
--- a/docs/video-canvas-audio-director-2026-09-17.md
+++ b/docs/video-canvas-audio-director-2026-09-17.md
@ -0,0 +1,40 @@
+# 阿里云音频与导演台扩展实施记录
+
+日期：2026-09-17。此记录接续 video-canvas-task-repairs-2026-09-17.md。
+
+## 已完成代码
+
+- 语音/音乐节点从仅上传改为官方模型选择、文本、音色、语言、格式、指令和歌曲参数。控件严格跟随所选模型能力，千问3 TTS 不发送其不支持的数值音调/速度参数。
+- 阿里云 HTTP 适配：qwen3-tts-flash、qwen3-tts-instruct-flash、qwen-audio-3.0-tts-flash/plus、fun-music-v1。北京限定模型不能错用国际服务地址；使用原服务商密钥，不给浏览器返回密钥。
+- 音频加入持久任务队列。先记录供应商原文件，再下载校验 WAV/MP3 并归档；转存失败只恢复下载，网络受理状态不明不会自动再次生成。
+- 导演台 AI 识图入口：调用当前视觉理解模型，把参考图转为人物/立方体/球体和机位组成的可编辑简化布景。结果独立保存，允许显式插入或替换原对象。不是精细纹理/网格重建，也没有凭空生成 GLB。
+- 对象轨道包含位置、旋转、缩放、可见性、启用开关、线性/缓入缓出插值、逐帧时间编辑/删除、自动记录、秒/30fps 显示单位、时间轴缩放。旋转按最短弧插值；对象删除清理轨道；修改总时长同步缩放轨道。
+- 相机与对象共用播放头和导出时间；跟随相机采样运动后的目标对象。纯对象动画也可预览和导出。场景骨骼动作序列、声音轨道、跨机位自动切镜尚未实现。
+- 修正浏览器 crypto.randomUUID 失去调用上下文导致记录帧/插入布景报 Illegal invocation；修正视频编码器立即停止可能丢掉末帧，预热首帧并留出末帧呈现时间。
+
+## 48 服务器配置
+
+已使用 `scripts/configure-aliyun-canvas-audio.cjs --apply` 写入现有阿里云官方北京服务商，保留其他模型和全局默认设置；变更前模型记录备份在应用目录 `.model-config-backups/`（不含服务商密钥）。
+
+- qwen3-tts-instruct-flash：启用，音频页面优先顺序 900。
+- qwen3-tts-flash：启用，顺序 901。
+- fun-music-v1：已建配置，禁用，标记 activationRequired，等待账号邀测权限确认。
+- Qwen Audio 3.0：适配已写，当前 Key 的只读 models 目录未列出，未盲目启用。
+- 千问3.8既有视觉理解配置保留。
+
+只读调用官方模型目录返回 HTTP 200，包含千问3 TTS/指令版；未发起真实付费语音/音乐/识图生成。模型目录可见性不能代替生成效果验收。
+
+[千问 TTS 官方文档](https://help.aliyun.com/zh/model-studio/qwen-tts-api)；[新音频 HTTP 文档](https://help.aliyun.com/zh/model-studio/cosyvoice-tts-http-api)；[Fun-Music 官方文档](https://help.aliyun.com/zh/model-studio/fun-music-api)（邀测、北京地域、临时文件 24 小时）。
+
+## 验证
+
+- helpers 全套 475 项通过；追加总时长/删除对象轨道用例后导演台 24 项通过，合计覆盖 476 项。
+- TypeScript `tsc --noEmit --incremental false` 通过。
+- `scripts/qa/video-canvas-audio.cjs` 使用实际服务适配器和本地替身：请求参数、归档前检查点、存储失败后零重复 POST、链接过期、服务商错误、租约丢失均通过。
+- `scripts/qa/video-canvas-task-db.cjs` 实际临时 PostgreSQL 验证通过：并发幂等、独占领取、恢复与失效 worker 隔离。
+- 本地浏览器调用实际 React 组件和 Three.js：音频能力控件、未开通音乐拦截、对象播放头/关键帧编辑/帧单位、独立对象动画视频导出、重建结果插入/替换通过，无浏览器运行错误。模型与上传返回使用本地替身。
+- 真实浏览器录制文件经 ffprobe 确认为 VP9、1280×720；设置 2 秒，文件约 1.974 秒。提取首尾帧确认对象实际移动。MediaRecorder 属于实时录制，受浏览器帧调度影响，不能宣称离线逐帧精确编码。
+
+参考站及开源自动剪辑算法详见 `video-canvas-smart-edit-research-2026-09-17.md`。该部分为已完成研究和实施方案，尚未接入完整自动混剪服务。
+
+部署状态：本次生产环境仅更新上述模型配置；应用代码在开发分支完成，未发布到生产应用。
--- a/docs/video-canvas-smart-edit-research-2026-09-17.md
+++ b/docs/video-canvas-smart-edit-research-2026-09-17.md
@ -0,0 +1,65 @@
+# 智能剪辑：参考站实测与开源算法选型
+
+检查日期：2026-09-17。本文区分实际观察、开源代码事实、建议方案与尚未实现的功能。
+
+## 参考站已观察
+
+[LibTV 画布](https://www.liblib.tv/canvas) 的智能剪辑入口有讲解视频、批量广告、口播视频、素材混剪四种预设。素材混剪预设连接七个素材节点，提示词包含卡点和突出产品，参数显示 16:9、1080P、20 秒。打开编辑器后，示例时间轴约 14 秒，含七个片段；这只是预设示例，不能当成自动生成成功的结果。没有点击收费生成。
+
+编辑器可见分割、向左/右裁剪、字幕轨道、转场、吸附、缩放、原声静音、全屏预览，导出到本地或画布。分割/裁剪会检查播放头位置与最短片段长度。界面没有公开它实际使用的后台选片或卡点算法，不能推断它使用了以下任何项目。
+
+## 已阅读的开源项目
+
+| 项目 | 可复用能力及实现 | 适用场景 | 许可证/接入边界 |
+| --- | --- | --- | --- |
+| [FunClip](https://github.com/modelscope/FunClip) | FunASR 生成转写与时间戳；把选定文字、说话人或 LLM 给出的区间映射回视频；裁剪后重排字幕时间 | 中文口播、讲解、访谈精选 | [源码 MIT](https://github.com/modelscope/FunClip/blob/main/LICENSE)，模型权重单独授权。Python/ASR 服务不能直接放进 Next 请求进程 |
+| [Auto-Editor](https://github.com/WyattBlue/auto-editor) | 按音量或运动量标记保留区间，可组合条件，前后 margin 保留呼吸和语句边界 | 去停顿、去静音粗剪 | [Unlicense](https://github.com/WyattBlue/auto-editor/blob/master/LICENSE)；低音量不是无价值内容，不能直接当产品亮点评分 |
+| [PySceneDetect](https://github.com/Breakthrough/PySceneDetect) | ContentDetector 计算相邻帧变化；AdaptiveDetector 将目标帧分数与周围帧均值比较，同时检查最小内容变化和最短镜头 | 多素材拆镜头、混剪候选片段 | [BSD-3-Clause](https://github.com/Breakthrough/PySceneDetect/blob/main/LICENSE)；负责边界检测，不负责语义或商品识别 |
+| [librosa](https://github.com/librosa/librosa) | onset strength → 节奏估计 → 动态规划选节拍；综合局部强度与节拍间隔一致性 | 背景音乐卡点 | [ISC](https://github.com/librosa/librosa/blob/main/LICENSE.md)；弱节奏/静音可能无有效节拍，必须显式回退 |
+| [FFmpeg](https://ffmpeg.org/ffmpeg-filters.html) | trim/atrim、setpts/asetpts、concat、xfade/acrossfade、字幕渲染、响度处理 | 最终确定性合成 | 使用部署环境既有构建；具体许可依构建组件，不把所有构建概括成同一许可 |
+
+阅读位置：FunClip `funclip/videoclipper.py` 的 `recog`、`clip`、`video_clip`；PySceneDetect `scenedetect/detectors/adaptive_detector.py` 的 `process_frame`；librosa `librosa/beat.py` 的 `__beat_tracker`、`__beat_track_dp`。这些代码已下载到本地研究目录查看，没有安装来源不明的脚本或启动重模型服务。
+
+FunClip 的关键点是可靠的时间戳映射，而不是让语言模型随意猜视频秒数。它既接收文字匹配，也接收显式时间区间，再重算片段字幕。不能把“ASR 文字剪辑”直接等同于“无对白产品素材混剪”。
+
+PySceneDetect 的自适应比值为目标变化分数除以邻域平均变化分数，另加内容阈值和最短镜头限制。这能减轻连续快速运镜带来的误切，但闪光和缓慢转场仍需结合素材调阈值。[算法实现](https://github.com/Breakthrough/PySceneDetect/blob/main/scenedetect/detectors/adaptive_detector.py)
+
+librosa 的节拍跟踪先形成起音强度曲线，估计速度，再用动态规划寻找与速度相容的强起音序列，不是简单按固定秒数平均切。低置信度时保留普通剪辑并标明未检测到稳定节拍。[算法文档](https://librosa.org/doc/0.10.2/generated/librosa.beat.beat_track.html)
+
+## 与现有系统的接入方案（尚未上线）
+
+优先组合 FunClip 的时间戳选片思路、PySceneDetect、librosa 和现有 FFmpeg；千问3.8继续负责素材语义、脚本要求和候选镜头排序。无需为了剪辑更换现有生图/视频模型。
+
+1. 预处理：现有受限下载器取得素材；ffprobe 读真实时长/FPS/音轨；以素材 hash 缓存低清代理、镜头边界、ASR 和节拍，避免反复解析。
+2. 候选分段：有对白时按 ASR 句子和说话人形成候选；无对白时按场景边界形成候选。去静音只作可选粗剪，保留前后余量，不切断语词。
+3. 语义选片：把候选 ID、真实开始/结束、抽帧与转写传给现有千问3.8，只允许模型返回候选 ID、用途和排序。服务端校验所有 ID 和区间；无素材时不能虚构成片。
+4. 卡点：分析用户选定/生成的音乐，输出真实 beat 秒数。根据最低镜头时长与素材余量，选每 2/4 拍换镜头，平衡节奏、语义连续性和目标总时长。口播优先语句边界，音乐卡点不能强行截断对白。
+5. 形成可编辑 EDL：记录素材 ID、sourceIn/sourceOut、timelineStart、speed、音量、字幕、转场。前端预览、时间轴修改和后台导出使用同一份 EDL。用整数微秒或有理数时间避免浮点累计漂移，显示时转换为帧。
+6. 合成：统一尺寸/FPS/像素格式/采样率；每段 trim 后时间归零。转场重叠计入总时长；无音轨片段补静音以保持 A/V 对齐；FFmpeg 执行显式参数数组，进程限时限并发，上传归档。
+7. 持久任务：分为解析、选片、编辑计划、渲染、转存，各步保存结果和输入版本。刷新或断线恢复同一任务，编辑后旧任务不能覆盖新版本；失败显示具体阶段。
+
+示例 EDL 合同：
+
+```json
+{
+  "version": 1,
+  "timebase": "microseconds",
+  "canvas": {"width": 1920, "height": 1080, "fps": 30},
+  "clips": [{"id":"clip-1","assetId":"asset-a","sourceIn":1000000,"sourceOut":3000000,"timelineStart":0,"speed":1,"volume":1}],
+  "music": {"assetId":"music-a","start":0,"gain":0.2},
+  "subtitles": [],
+  "transitions": []
+}
+```
+
+## 对照记录与剩余工作
+
+| 能力 | 当前项目状态 | 后续验收条件 |
+| --- | --- | --- |
+| 单视频裁剪、抽帧、音轨提取、慢放 | 现有真实 FFmpeg 功能 | 已有测试继续保持 |
+| 四类智能剪辑预设的自动成片 | 本次完成参考站与开源源码研究，未接入完整剪辑服务 | 上传不同素材后产生可编辑 EDL 与真实成片 |
+| 去静音/ASR 语义选片 | 选定开源候选，尚未部署 | 验证中文、低音量、无音轨、多人语音 |
+| 音乐卡点 | 已明确 librosa 算法与回退策略，尚未接入 | 合成脉冲音乐和真实音乐的切点误差可量化 |
+| 多素材时间轴、字幕、转场 | 仍需 EDL 编辑器和合成执行器；不把单片裁剪宣称为智能混剪 | 预览/导出逐片一致，跨转场音画同步 |
+
+48 服务器此前反馈过内存高。部署 ASR 时应先核对实际可用资源，再使用独立受限 worker；不要在每个 Web 请求内加载一份模型。此处是实施建议，本次没有启动新的常驻模型服务。
--- a/lib/video-canvas/audio-contract.ts
+++ b/lib/video-canvas/audio-contract.ts
@ -0,0 +1,197 @@
+import { z } from "zod";
+const qwenVoices = [
+  { id: "Cherry", name: "芊悦" },
+  { id: "Serena", name: "苏瑶" },
+];
+export const audioModelContracts = {
+  "qwen3-tts-instruct-flash": {
+    kind: "tts",
+    label: "千问3 TTS 指令版",
+    voices: qwenVoices,
+    instructions: true,
+    controls: false,
+    formats: ["wav"],
+  },
+  "qwen3-tts-flash": {
+    kind: "tts",
+    label: "千问3 TTS",
+    voices: [...qwenVoices, { id: "Ryan", name: "甜茶" }, { id: "Jennifer", name: "詹妮弗" }],
+    instructions: false,
+    controls: false,
+    formats: ["wav"],
+  },
+  "qwen-audio-3.0-tts-flash": {
+    kind: "tts",
+    label: "千问 Audio 3.0 Flash",
+    voices: [
+      { id: "longanhuan_v3.6", name: "龙安欢" },
+      { id: "longanfengyue", name: "龙安风悦" },
+      { id: "longanyuanfei", name: "龙安元妃" },
+      { id: "longanlingxi", name: "龙安灵希" },
+    ],
+    instructions: true,
+    controls: true,
+    formats: ["mp3", "wav"],
+  },
+  "qwen-audio-3.0-tts-plus": {
+    kind: "tts",
+    label: "千问 Audio 3.0 Plus",
+    voices: [
+      { id: "longanlingxin", name: "龙安灵心" },
+      { id: "longanlufeng", name: "龙安鲁风" },
+    ],
+    instructions: true,
+    controls: true,
+    formats: ["mp3", "wav"],
+  },
+  "fun-music-v1": {
+    kind: "music",
+    label: "Fun-Music 音乐",
+    voices: [],
+    instructions: false,
+    controls: false,
+    formats: ["mp3", "wav"],
+  },
+} satisfies Record<
+  string,
+  {
+    kind: string;
+    label: string;
+    voices: { id: string; name: string }[];
+    instructions: boolean;
+    controls: boolean;
+    formats: string[];
+  }
+>;
+export type AudioModelId = keyof typeof audioModelContracts;
+export const audioRequestSchema = z.object({
+  mode: z.enum(["tts", "music"]),
+  modelRef: z.string().min(1),
+  text: z.string().trim().max(2000).default(""),
+  lyrics: z.string().trim().max(2000).default(""),
+  voice: z.string().max(100).optional(),
+  language: z
+    .enum([
+      "Auto",
+      "Chinese",
+      "English",
+      "German",
+      "Italian",
+      "Portuguese",
+      "Spanish",
+      "Japanese",
+      "Korean",
+      "French",
+      "Russian",
+    ])
+    .default("Chinese"),
+  instructions: z.string().trim().max(1200).default(""),
+  format: z.enum(["mp3", "wav"]).default("wav"),
+  instrumental: z.boolean().default(true),
+  gender: z.enum(["male", "female"]).default("female"),
+  rate: z.number().min(0.5).max(2).default(1),
+  pitch: z.number().min(0.5).max(2).default(1),
+  volume: z.number().int().min(0).max(100).default(50),
+});
+export type AudioRequest = z.infer<typeof audioRequestSchema>;
+export function audioContract(id: string) {
+  return Object.hasOwn(audioModelContracts, id)
+    ? audioModelContracts[id as AudioModelId]
+    : undefined;
+}
+export function officialAudioBase(base: string, modelId?: string) {
+  const url = new URL(base);
+  if (
+    url.protocol !== "https:" ||
+    url.username ||
+    url.password ||
+    !/^(?:dashscope(?:-intl|-us)?\.aliyuncs\.com|[a-zA-Z0-9-]+\.(?:cn-beijing|ap-southeast-1)\.maas\.aliyuncs\.com)$/.test(
+      url.hostname,
+    )
+  )
+    throw Error("请选择阿里云官方音频服务线路");
+  if (
+    (modelId === "fun-music-v1" || modelId?.startsWith("qwen-audio-")) &&
+    url.hostname !== "dashscope.aliyuncs.com" &&
+    !url.hostname.endsWith(".cn-beijing.maas.aliyuncs.com")
+  )
+    throw Error("所选模型仅支持阿里云北京地域");
+  return url.origin;
+}
+export function buildAudioRequest(value: unknown, modelId: string) {
+  const p = audioRequestSchema.parse(value),
+    contract = audioContract(modelId);
+  if (!contract || contract.kind !== p.mode) throw Error("音频模式与模型不匹配");
+  if (!contract.formats.includes(p.format)) throw Error("所选模型不支持这个音频格式");
+  if (p.mode === "music") {
+    if ((p.instrumental && !p.text) || (!p.text && !p.lyrics)) throw Error("请输入音乐描述或歌词");
+    if (
+      !p.instrumental &&
+      p.lyrics &&
+      (p.lyrics.length < 5 || (/[\u3400-\u9fff]/.test(p.lyrics) && p.lyrics.length > 350))
+    )
+      throw Error("中文歌词需 5–350 字，英文歌词需 5–2000 字");
+    return {
+      path: "/api/v1/services/audio/music/generation",
+      body: {
+        model: modelId,
+        input: {
+          ...(p.text ? { prompt: p.text } : {}),
+          ...(!p.instrumental && p.lyrics ? { lyrics: p.lyrics } : {}),
+          is_instrumental: p.instrumental,
+          ...(!p.instrumental ? { gender: p.gender } : {}),
+          format: p.format,
+        },
+      },
+    };
+  }
+  if (!p.text || p.text.length > 600) throw Error("配音文本需 1–600 字");
+  if (!contract.voices.some((v) => v.id === p.voice)) throw Error("请选择当前模型支持的音色");
+  if (!contract.instructions && p.instructions) throw Error("所选模型不支持风格指令");
+  if (!contract.controls && (p.rate !== 1 || p.pitch !== 1 || p.volume !== 50))
+    throw Error("所选模型不支持数值语速、音调和音量参数，可改用风格指令");
+  const modern = modelId.startsWith("qwen-audio-");
+  if (modern && !["Chinese", "English", "Auto"].includes(p.language))
+    throw Error("当前系统音色支持中文和英文");
+  return {
+    path: modern
+      ? "/api/v1/services/audio/tts/SpeechSynthesizer"
+      : "/api/v1/services/aigc/multimodal-generation/generation",
+    body: {
+      model: modelId,
+      input: {
+        text: p.text,
+        voice: p.voice,
+        ...(modern
+          ? {
+              format: p.format,
+              rate: p.rate,
+              pitch: p.pitch,
+              volume: p.volume,
+              ...(p.instructions ? { instruction: p.instructions } : {}),
+              ...(p.language !== "Auto"
+                ? { language_hints: [p.language === "Chinese" ? "zh" : "en"] }
+                : {}),
+            }
+          : {
+              language_type: p.language,
+              ...(p.instructions
+                ? { instructions: p.instructions, optimize_instructions: true }
+                : {}),
+            }),
+      },
+    },
+  };
+}
+export const audioCheckpointSchema = z.object({
+  operation: z.literal("audio"),
+  modelRef: z.string(),
+  modelId: z.string(),
+  url: z
+    .string()
+    .url()
+    .refine((v) => /^https?:\/\//i.test(v)),
+  format: z.enum(["mp3", "wav"]),
+  requestId: z.string().optional(),
+  expiresAt: z.number().optional(),
+});
--- a/lib/video-canvas/audio-generation.ts
+++ b/lib/video-canvas/audio-generation.ts
@ -0,0 +1,132 @@
+import "server-only";
+import { mkdtemp, readFile, rm } from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { getEnabledAiModelsWithProvider } from "@/lib/db/queries";
+import { uploadFile } from "@/lib/storage/service";
+import { downloadCanvasMedia } from "./download-media";
+import {
+  audioContract,
+  audioCheckpointSchema,
+  audioRequestSchema,
+  buildAudioRequest,
+  officialAudioBase,
+} from "./audio-contract";
+import { ProviderTaskUncertainError } from "./video-task-lifecycle";
+export async function listCanvasAudioModels() {
+  return (await getEnabledAiModelsWithProvider()).filter(({ model, provider }) => {
+    try {
+      return (
+        provider.type === "alibaba" &&
+        !!audioContract(model.modelId) &&
+        !!officialAudioBase(provider.baseUrl, model.modelId)
+      );
+    } catch {
+      return false;
+    }
+  });
+}
+export async function executeCanvasAudio(
+  params: unknown,
+  lifecycle: {
+    resume?: unknown;
+    assertActive: () => Promise<void>;
+    checkpoint: (value: Record<string, unknown>) => Promise<void>;
+  },
+) {
+  const input = audioRequestSchema.parse(params);
+  let checkpoint = lifecycle.resume ? audioCheckpointSchema.parse(lifecycle.resume) : undefined;
+  if (checkpoint && checkpoint.modelRef !== input.modelRef) throw Error("音频记录与原模型不匹配");
+  if (!checkpoint) {
+    const item = (await listCanvasAudioModels()).find((x) => x.model.id === input.modelRef);
+    if (!item) throw Error("音频模型未启用，请先在阿里云和模型设置中开通");
+    const request = buildAudioRequest(input, item.model.modelId);
+    await lifecycle.assertActive();
+    let response: Response;
+    try {
+      response = await fetch(
+        officialAudioBase(item.provider.baseUrl, item.model.modelId) + request.path,
+        {
+          method: "POST",
+          headers: {
+            Authorization: `Bearer ${item.provider.apiKey}`,
+            "Content-Type": "application/json",
+          },
+          body: JSON.stringify(request.body),
+          signal: AbortSignal.timeout(input.mode === "music" ? 600000 : 180000),
+        },
+      );
+    } catch {
+      throw new ProviderTaskUncertainError(
+        "音频请求连接中断，无法确认是否已受理；请核对服务商记录，系统不会自动重发",
+      );
+    }
+    let data: any;
+    try {
+      data = await response.json();
+    } catch {
+      throw new ProviderTaskUncertainError("音频响应无法解析，请核对服务商记录，系统不会自动重发");
+    }
+    if (!response.ok)
+      throw Error(
+        `阿里云音频请求失败（${response.status}，${String(data.code || data.error?.code || "UnknownError").slice(0, 120)}）${input.mode === "music" ? "，请确认 Fun-Music 邀测权限已开通" : ""}`,
+      );
+    const url = data.output?.audio?.url;
+    if (typeof url !== "string" || !/^https?:\/\//i.test(url))
+      throw new ProviderTaskUncertainError("音频服务未返回有效文件地址，请核对原请求，勿重复生成");
+    checkpoint = {
+      operation: "audio",
+      url,
+      format: input.format,
+      modelRef: item.model.id,
+      modelId: item.model.modelId,
+      requestId: data.request_id,
+      expiresAt:
+        typeof data.output.audio.expires_at === "number" ? data.output.audio.expires_at : undefined,
+    };
+    try {
+      await lifecycle.assertActive();
+      await lifecycle.checkpoint({ upstreamAudio: checkpoint });
+    } catch {
+      throw new ProviderTaskUncertainError(
+        "音频已生成，但保存任务记录失败；请核对原请求，系统不会自动重发",
+      );
+    }
+  }
+  if (checkpoint.expiresAt && checkpoint.expiresAt * 1000 < Date.now())
+    throw Error("阿里云音频临时地址已过期，无法继续转存");
+  await lifecycle.assertActive();
+  const dir = await mkdtemp(path.join(os.tmpdir(), "canvas-audio-"));
+  try {
+    const file = path.join(dir, "audio." + checkpoint.format);
+    await downloadCanvasMedia(checkpoint.url, file);
+    const bytes = await readFile(file);
+    if (bytes.length < 12) throw Error("音频内容为空");
+    const wav =
+      bytes.toString("ascii", 0, 4) === "RIFF" && bytes.toString("ascii", 8, 12) === "WAVE";
+    const mp3 =
+      bytes.toString("ascii", 0, 3) === "ID3" || (bytes[0] === 0xff && (bytes[1] & 0xe0) === 0xe0);
+    if (!(checkpoint.format === "wav" ? wav : mp3)) throw Error("返回文件不是有效音频格式");
+    await lifecycle.assertActive();
+    const saved = await uploadFile(
+      `audio-${crypto.randomUUID()}.${checkpoint.format}`,
+      Uint8Array.from(bytes).buffer,
+    );
+    return {
+      url: saved.url,
+      modelRef: checkpoint.modelRef,
+      modelId: checkpoint.modelId,
+      metadata: {
+        storageStatus: "archived",
+        format: checkpoint.format,
+        requestId: checkpoint.requestId,
+      },
+    };
+  } catch (error) {
+    throw new ProviderTaskUncertainError(
+      `音频已生成，转存暂未完成，将继续转存原文件：${error instanceof Error ? error.message : "存储失败"}`,
+    );
+  } finally {
+    await rm(dir, { recursive: true, force: true });
+  }
+}
--- a/lib/video-canvas/director-extensions.ts
+++ b/lib/video-canvas/director-extensions.ts
@ -1,4 +1,10 @@
-import { defaultDirectorMotion, type DirectorPose, type DirectorMotion } from "./director-motion";
+import { resizeObjectTracks } from "./director-object-tracks";
+import {
+  defaultDirectorMotion,
+  resizeDirectorMotion,
+  type DirectorPose,
+  type DirectorMotion,
+} from "./director-motion";
 import type {
  DirectorScene,
  DirectorCamera,
@ -14,11 +20,27 @@ export function commitDirectorScene(previous: DirectorScene, next: DirectorScene
  const motion =
    previous.activeCameraId === next.activeCameraId
      ? next.motion
-      : motionTracks[next.activeCameraId] || defaultDirectorMotion();
+      : motionTracks[next.activeCameraId] || {
+          ...defaultDirectorMotion(),
+          duration: previous.motion.duration,
+        };
  motionTracks[next.activeCameraId] = motion;
  for (const id of Object.keys(motionTracks))
    if (!next.cameras.some((c) => c.id === id)) delete motionTracks[id];
-  return { ...next, motion, motionTracks };
+  let objectTracks = next.objectTracks;
+  if (
+    previous.activeCameraId === next.activeCameraId &&
+    previous.motion.duration !== next.motion.duration
+  ) {
+    for (const id of Object.keys(motionTracks))
+      if (id !== next.activeCameraId)
+        motionTracks[id] = resizeDirectorMotion(motionTracks[id], next.motion.duration);
+    objectTracks = resizeObjectTracks(objectTracks, previous.motion.duration, next.motion.duration);
+  }
+  objectTracks = Object.fromEntries(
+    Object.entries(objectTracks || {}).filter(([id]) => next.objects.some((o) => o.id === id)),
+  );
+  return { ...next, motion, motionTracks, objectTracks };
 }
 export function resolveDirectorCamera(scene: DirectorScene, pose?: DirectorPose): DirectorCamera {
  const camera = scene.cameras.find((c) => c.id === scene.activeCameraId) || scene.cameras[0];
--- a/lib/video-canvas/director-object-tracks.ts
+++ b/lib/video-canvas/director-object-tracks.ts
@ -0,0 +1,147 @@
+import type { DirectorScene, DirectorObject, DirectorVector } from "./director-scene";
+export type DirectorObjectFrame = {
+  id: string;
+  time: number;
+  position: DirectorVector;
+  rotation: DirectorVector;
+  scale: DirectorVector;
+  visible: boolean;
+};
+export type DirectorObjectTrack = {
+  muted: boolean;
+  easing: "linear" | "smooth";
+  keyframes: DirectorObjectFrame[];
+};
+export function readObjectTracks(
+  value: unknown,
+  objects: DirectorObject[],
+  duration: number,
+): Record<string, DirectorObjectTrack> {
+  if (!value || typeof value !== "object" || Array.isArray(value)) return {};
+  const vector = (v: any, positive = false) =>
+    Array.isArray(v) &&
+    v.length === 3 &&
+    v.every(
+      (n: any) =>
+        typeof n === "number" && Number.isFinite(n) && Math.abs(n) <= 10000 && (!positive || n > 0),
+    );
+  return Object.fromEntries(
+    Object.entries(value)
+      .filter(([id]) => objects.some((o) => o.id === id))
+      .map(([id, raw]) => {
+        const t = raw as any,
+          seen = new Set<number>(),
+          ids = new Set<string>();
+        const keyframes = (Array.isArray(t?.keyframes) ? t.keyframes : [])
+          .filter((k: any) => {
+            if (
+              !k ||
+              typeof k.id !== "string" ||
+              ids.has(k.id) ||
+              typeof k.time !== "number" ||
+              !Number.isFinite(k.time) ||
+              k.time < 0 ||
+              k.time > duration ||
+              seen.has(k.time) ||
+              !vector(k.position) ||
+              !vector(k.rotation) ||
+              !vector(k.scale, true)
+            )
+              return false;
+            seen.add(k.time);
+            ids.add(k.id);
+            return true;
+          })
+          .slice(0, 100)
+          .map((k: any) => ({
+            id: k.id,
+            time: k.time,
+            position: [...k.position],
+            rotation: [...k.rotation],
+            scale: [...k.scale],
+            visible: k.visible !== false,
+          }))
+          .sort((a: any, b: any) => a.time - b.time);
+        return [
+          id,
+          {
+            muted: t?.muted === true,
+            easing: t?.easing === "smooth" ? "smooth" : "linear",
+            keyframes,
+          },
+        ];
+      }),
+  );
+}
+export function sampleDirectorObjects(scene: DirectorScene, time: number | null): DirectorObject[] {
+  if (time === null) return scene.objects;
+  return scene.objects.map((object) => {
+    const track = scene.objectTracks?.[object.id];
+    if (!track || track.muted || !track.keyframes.length) return object;
+    const keys = track.keyframes,
+      t = Math.max(0, Math.min(scene.motion.duration, time)),
+      right = keys.findIndex((k) => k.time > t);
+    const a = right < 0 ? keys[keys.length - 1] : keys[Math.max(0, right - 1)],
+      b = right < 0 ? a : keys[right];
+    let u = a === b ? 0 : Math.max(0, Math.min(1, (t - a.time) / (b.time - a.time)));
+    if (track.easing === "smooth") u = u * u * (3 - 2 * u);
+    const vector = (x: DirectorVector, y: DirectorVector, rotation = false): DirectorVector =>
+      x.map((n, i) => {
+        let delta = y[i] - n;
+        if (rotation) delta = Math.atan2(Math.sin(delta), Math.cos(delta));
+        return n + delta * u;
+      }) as DirectorVector;
+    return {
+      ...object,
+      position: vector(a.position, b.position),
+      rotation: vector(a.rotation, b.rotation, true),
+      scale: vector(a.scale, b.scale),
+      visible: a.visible,
+    };
+  });
+}
+export function recordObjectFrames(
+  scene: DirectorScene,
+  ids: string[],
+  time: number,
+  id: () => string,
+): DirectorScene {
+  if (!Number.isFinite(time) || time < 0 || time > scene.motion.duration)
+    throw Error("关键帧时间超出时间轴");
+  const tracks = { ...scene.objectTracks };
+  for (const object of scene.objects.filter((o) => ids.includes(o.id))) {
+    const track = tracks[object.id] || { muted: false, easing: "linear" as const, keyframes: [] };
+    const existing = track.keyframes.find((k) => Math.abs(k.time - time) < 0.01);
+    if (!existing && track.keyframes.length >= 100) throw Error("每个对象最多 100 个关键帧");
+    const frame = {
+      id: existing?.id || id(),
+      time: existing?.time ?? Math.round(time * 100) / 100,
+      position: [...object.position] as DirectorVector,
+      rotation: [...object.rotation] as DirectorVector,
+      scale: [...object.scale] as DirectorVector,
+      visible: object.visible !== false,
+    };
+    tracks[object.id] = {
+      ...track,
+      keyframes: [...track.keyframes.filter((k) => k.id !== frame.id), frame].sort(
+        (a, b) => a.time - b.time,
+      ),
+    };
+  }
+  return { ...scene, objectTracks: tracks };
+}
+export function resizeObjectTracks(
+  tracks: Record<string, DirectorObjectTrack> | undefined,
+  oldDuration: number,
+  newDuration: number,
+) {
+  return Object.fromEntries(
+    Object.entries(tracks || {}).map(([id, t]) => [
+      id,
+      {
+        ...t,
+        keyframes: t.keyframes.map((k) => ({ ...k, time: (k.time * newDuration) / oldDuration })),
+      },
+    ]),
+  );
+}
--- a/lib/video-canvas/director-reconstruction.ts
+++ b/lib/video-canvas/director-reconstruction.ts
@ -0,0 +1,69 @@
+import { z } from "zod";
+import { defaultDirectorScene, type DirectorScene } from "./director-scene";
+import { validDirectorPose } from "./director-motion";
+const vector = z.tuple([
+  z.number().finite().min(-100).max(100),
+  z.number().finite().min(-100).max(100),
+  z.number().finite().min(-100).max(100),
+]);
+const scale = z.tuple([
+  z.number().min(0.05).max(30),
+  z.number().min(0.05).max(30),
+  z.number().min(0.05).max(30),
+]);
+const reconstructionSchema = z.object({
+  summary: z.string().min(1).max(2000),
+  objects: z
+    .array(
+      z.object({
+        name: z.string().min(1).max(80),
+        kind: z.enum(["actor", "box", "sphere"]),
+        position: vector,
+        rotation: vector,
+        scale,
+        color: z.string().regex(/^#[0-9a-f]{6}$/i),
+      }),
+    )
+    .min(1)
+    .max(30),
+  camera: z
+    .object({ position: vector, target: vector, fov: z.number().min(15).max(90) })
+    .refine(validDirectorPose, "相机不能与注视点重合"),
+});
+export function reconstructionPrompt(prompt: string) {
+  return `根据参考图重建可编辑的3D布景（灰盒），不是输出图片或GLB。只输出JSON，字段：summary、objects、camera。只重建看得见的主要人物和物体，忽略细碎纹理，不编造遮挡后的形状。以米为近似单位，Y轴向上，地面y=0；对象局部几何 actor 约2米高，box边长1米，sphere直径1米，scale必须据此设置；对象中心/脚底位置遵循 actor脚底、box和sphere中心。rotation是弧度。所有坐标在±100内，scale在0.05到30内。objects最多30项，每项{name,kind:"actor"|"box"|"sphere",position:[x,y,z],rotation:[rx,ry,rz],scale:[sx,sy,sz],color:"#RRGGBB"}。camera={position:[x,y,z],target:[x,y,z],fov:15到90}，还原参考图的构图。summary说明不确定的比例与缺失信息。用户布景要求：${prompt || "尽量保留原图主体布局与机位"}`;
+}
+export function parseReconstructedScene(value: unknown): {
+  summary: string;
+  directorScene: DirectorScene;
+} {
+  const parsed = reconstructionSchema.parse(value),
+    base = defaultDirectorScene();
+  return {
+    summary: parsed.summary,
+    directorScene: {
+      ...base,
+      objects: parsed.objects.map((o, i) => ({ ...o, id: `reconstructed-${i}` })),
+      cameras: [{ ...parsed.camera, id: "reconstructed-camera", name: "参考图机位" }],
+      activeCameraId: "reconstructed-camera",
+    },
+  };
+}
+export function insertReconstructedScene(
+  current: DirectorScene,
+  generated: DirectorScene,
+  replace: boolean,
+  id: () => string,
+): DirectorScene {
+  const objects = generated.objects.map((o) => ({ ...o, id: id() }));
+  if ((replace ? 0 : current.objects.length) + objects.length > 200)
+    throw Error("场景最多支持 200 个对象");
+  const camera = { ...generated.cameras[0], id: id(), followObjectId: undefined };
+  return {
+    ...current,
+    objects: replace ? objects : [...current.objects, ...objects],
+    cameras: [...current.cameras, camera],
+    activeCameraId: camera.id,
+    ...(replace ? { objectTracks: {} } : {}),
+  };
+}
--- a/lib/video-canvas/director-recording.ts
+++ b/lib/video-canvas/director-recording.ts
@ -21,6 +21,30 @@ export async function recordDirectorVideo({
  if (!mimeType || !canvas.captureStream)
    throw Error("当前浏览器不支持视频录制，请更换支持视频录制的浏览器");
  if (signal.aborted || document.hidden) throw Error("请保持当前页面可见后导出");
+  // Let WebGL present the first frame before opening the encoder stream.
+  render(0);
+  await new Promise<void>((resolve, reject) => {
+    let frame = 0;
+    const finish = (error?: Error) => {
+      cancelAnimationFrame(frame);
+      clearTimeout(timeout);
+      signal.removeEventListener("abort", cancel);
+      document.removeEventListener("visibilitychange", hidden);
+      if (error) reject(error);
+      else resolve();
+    };
+    const cancel = () => finish(Error("已取消轨迹视频导出"));
+    const hidden = () => {
+      if (document.hidden) finish(Error("导出时切换了页面，请回到当前页面重新导出"));
+    };
+    const timeout = setTimeout(() => finish(Error("画面准备超时，请保持页面可见后重试")), 2000);
+    signal.addEventListener("abort", cancel, { once: true });
+    document.addEventListener("visibilitychange", hidden);
+    frame = requestAnimationFrame(() => {
+      frame = requestAnimationFrame(() => finish());
+    });
+  });
+  if (signal.aborted || document.hidden) throw Error("请保持当前页面可见后导出");
  const stream = canvas.captureStream(30);
  try {
    return await new Promise<Blob>((resolve, reject) => {
@ -78,7 +102,12 @@ export async function recordDirectorVideo({
          stop(e instanceof Error ? e : Error("轨迹渲染失败"));
          return;
        }
-        if (t >= duration) stop();
+        // MediaRecorder consumes the canvas asynchronously. Stopping in the rendering
+        // callback drops the final pose; allow two presentation frames to flush it.
+        if (t >= duration)
+          frame = requestAnimationFrame(() => {
+            frame = requestAnimationFrame(() => stop());
+          });
        else frame = requestAnimationFrame(tick);
      };
      try {
--- a/lib/video-canvas/director-scene.ts
+++ b/lib/video-canvas/director-scene.ts
@ -1,3 +1,4 @@
+import { readObjectTracks, type DirectorObjectTrack } from "./director-object-tracks";
 import { directorBodies } from "./director-extensions";
 import { directorPoses, type DirectorPoseName } from "./director-poses";
 import {
@ -45,6 +46,7 @@ export type DirectorScene = {
  snapshots: DirectorSnapshot[];
  motion: DirectorMotion;
  motionTracks?: Record<string, DirectorMotion>;
+  objectTracks?: Record<string, DirectorObjectTrack>;
  backgroundUrl?: string;
  backgroundHistory?: string[];
 };
@ -81,6 +83,7 @@ export function defaultDirectorScene(): DirectorScene {
    snapshots: [],
    motion: defaultDirectorMotion(),
    motionTracks: {},
+    objectTracks: {},
  };
 }
 export function readDirectorScene(value: unknown): DirectorScene {
@ -170,6 +173,7 @@ export function readDirectorScene(value: unknown): DirectorScene {
          ].slice(-20),
        }
      : {}),
+    objectTracks: readObjectTracks(v.objectTracks, objects, readDirectorMotion(v.motion).duration),
    motionTracks: Object.fromEntries(
      Object.entries(v.motionTracks || {})
        .filter(([id]) => validCameras.some((c) => c.id === id))
--- a/lib/video-canvas/task-contract.ts
+++ b/lib/video-canvas/task-contract.ts
@ -6,7 +6,7 @@ import { z } from "zod";
 export const canvasTaskRequestSchema = z.object({
  requestId: z.string().uuid(),
  projectId: z.string().uuid(),
-  kind: z.enum(["video", "image", "analysis", "image-operation"]),
+  kind: z.enum(["video", "image", "analysis", "image-operation", "audio"]),
  node: z
    .object({
      id: z.string().min(1).max(180),
@ -51,6 +51,14 @@ export function requireAnalysisResult(data: Record<string, any> | undefined, mod
  if (!data) throw Error("分析没有返回结果，原内容已保留");
  if (mode === "video-analysis" && !normalizeVideoAnalysisRows(data.analysisRows).length)
    throw Error("视频解析没有返回有效镜头，原内容已保留");
+  if (
+    mode === "director-reconstruct" &&
+    (!Array.isArray(data.directorScene?.objects) ||
+      !data.directorScene.objects.length ||
+      !Array.isArray(data.directorScene?.cameras) ||
+      !data.directorScene.cameras.length)
+  )
+    throw Error("识图重建没有返回有效布景，原场景已保留");
  const rows = data.analysisRows || data.shotIdeas;
  if (
    mode === "story-script" &&
@ -130,6 +138,15 @@ export function reconcileCanvasTasks<
        pendingLabel: "",
      });
      subtitle = "生成结果";
+      if (task.request.kind === "audio")
+        Object.assign(nextDraft, {
+          mediaUrl: data.url,
+          url: [data.url],
+          action: "AUDIO_RESOURCE",
+          mediaType: "audio",
+          audioMeta: data.metadata,
+          audioParams: task.request.params,
+        });
      if (task.request.kind === "video")
        Object.assign(nextDraft, { mediaUrl: data.url, url: [data.url], videoMeta: data.metadata });
      if (task.request.kind === "image" || task.request.kind === "image-operation") {
@ -156,6 +173,11 @@ export function reconcileCanvasTasks<
            ];
        }
      }
+      if (task.request.kind === "analysis" && task.request.params.mode === "director-reconstruct")
+        Object.assign(nextDraft, {
+          directorScene: data.directorScene,
+          reconstructionSummary: data.summary,
+        });
      if (task.request.kind === "analysis")
        Object.assign(nextDraft, {
          content: [data.raw || data.summary || data.agentPlan?.join("\n") || ""],
@ -210,7 +232,15 @@ export function reconcileCanvasTasks<
        ? {
            ...item,
            ...(task.status === "succeeded" && task.request.kind !== "analysis"
-              ? { kind: "media", inputType: task.request.kind === "video" ? "video" : "image" }
+              ? {
+                  kind: "media",
+                  inputType:
+                    task.request.kind === "video"
+                      ? "video"
+                      : task.request.kind === "audio"
+                        ? "audio"
+                        : "image",
+                }
              : {}),
            subtitle,
            draftValues: nextDraft,
--- a/lib/video-canvas/task-recovery.ts
+++ b/lib/video-canvas/task-recovery.ts
@ -1,3 +1,4 @@
+import { audioCheckpointSchema } from "./audio-contract";
 import { videoCheckpointSchema } from "./video-task-lifecycle";
 // Resume only server-recorded provider jobs. A lost request without an upstream ID is never resubmitted.
 export type RecoveryTask = {
@ -39,10 +40,17 @@ export function getVideoCheckpoint(task: Pick<RecoveryTask, "request" | "result"
    return null;
  return parsed.data;
 }
+export function getAudioCheckpoint(task: Pick<RecoveryTask, "request" | "result">) {
+  if (task.request.kind !== "audio") return null;
+  const parsed = audioCheckpointSchema.safeParse(task.result?.upstreamAudio);
+  return parsed.success && parsed.data.modelRef === task.request.params.modelRef
+    ? parsed.data
+    : null;
+}
 export function hasRecoverableCanvasTask(task: RecoveryTask, now = Date.now()) {
  const created = Date.parse(task.createdAt);
  return (
-    Boolean(getEnhanceCheckpoint(task) || getVideoCheckpoint(task)) &&
+    Boolean(getEnhanceCheckpoint(task) || getVideoCheckpoint(task) || getAudioCheckpoint(task)) &&
    Number.isFinite(created) &&
    now >= created &&
    now - created <= 86400000 &&
--- a/lib/video-canvas/task-worker.ts
+++ b/lib/video-canvas/task-worker.ts
@ -1,5 +1,6 @@
+import { executeCanvasAudio } from "./audio-generation";
 import { ProviderTaskUncertainError } from "./video-task-lifecycle";
-import { getEnhanceCheckpoint, getVideoCheckpoint } from "./task-recovery";
+import { getEnhanceCheckpoint, getVideoCheckpoint, getAudioCheckpoint } from "./task-recovery";
 import { requireStoryPromptResult } from "./story-task-attachment";
 import {
  executeImageOperation,
@ -37,9 +38,16 @@ async function execute(task: CanvasGenerationTask, userId: string) {
  await assertLease();
  const resume = getEnhanceCheckpoint(task);
  const videoResume = getVideoCheckpoint(task);
-  if ((task.attempts || 1) > 1 && !resume && !videoResume)
+  const audioResume = getAudioCheckpoint(task);
+  if ((task.attempts || 1) > 1 && !resume && !videoResume && !audioResume)
    throw Error("缺少上游任务编号，禁止重新提交");
  const params = request.params;
+  if (request.kind === "audio")
+    return executeCanvasAudio(params, {
+      resume: audioResume,
+      assertActive: assertLease,
+      checkpoint: (value) => checkpointCanvasGenerationTask(task.id, value, task.attempts),
+    });
  if (request.kind === "image-operation")
    return executeImageOperation(
      params,
--- a/lib/video-canvas/workflow-analysis-response.ts
+++ b/lib/video-canvas/workflow-analysis-response.ts
@ -1,3 +1,4 @@
+import { reconstructionPrompt, parseReconstructedScene } from "./director-reconstruction";
 import { normalizeVideoAnalysisRows } from "./video-story";
 import "server-only";
 import { normalizeStoryScriptRows } from "@/lib/video-canvas/story-script";
@ -22,7 +23,13 @@ const VIDEO_ANALYSIS_MAX_SIDE = 960;
 const VIDEO_ANALYSIS_TIMEOUT_MS = 15000;

 const requestSchema = z.object({
-  mode: z.enum(["video-analysis", "character-scene-extraction", "story-script", "agent-plan"]),
+  mode: z.enum([
+    "video-analysis",
+    "character-scene-extraction",
+    "story-script",
+    "agent-plan",
+    "director-reconstruct",
+  ]),
  prompt: z.string().trim().optional(),
  imageUrls: z.array(z.string().trim()).optional(),
  videoUrls: z.array(z.string().trim()).optional(),
@ -511,33 +518,35 @@ export async function executeWorkflowAnalysisResponse(body: unknown, signal?: Ab
    }

    const promptText =
-      parsed.data.mode === "video-analysis"
-        ? buildVideoAnalysisPrompt({
-            frameCount: extractedVideoFrames.length,
-            nodeTitle: normalizeText(parsed.data.nodeTitle),
-            prompt,
-            imageCount: imageUrls.length,
-            videoCount: videoUrls.length,
-          })
-        : parsed.data.mode === "character-scene-extraction"
-          ? buildCharacterScenePrompt({
+      parsed.data.mode === "director-reconstruct"
+        ? reconstructionPrompt(prompt)
+        : parsed.data.mode === "video-analysis"
+          ? buildVideoAnalysisPrompt({
+              frameCount: extractedVideoFrames.length,
              nodeTitle: normalizeText(parsed.data.nodeTitle),
              prompt,
              imageCount: imageUrls.length,
+              videoCount: videoUrls.length,
            })
-          : parsed.data.mode === "story-script"
-            ? buildStoryScriptPrompt({
+          : parsed.data.mode === "character-scene-extraction"
+            ? buildCharacterScenePrompt({
                nodeTitle: normalizeText(parsed.data.nodeTitle),
                prompt,
                imageCount: imageUrls.length,
-                videoCount: videoUrls.length,
              })
-            : buildAgentPlanPrompt({
-                nodeTitle: normalizeText(parsed.data.nodeTitle),
-                prompt,
-                imageCount: imageUrls.length,
-                videoCount: videoUrls.length,
-              });
+            : parsed.data.mode === "story-script"
+              ? buildStoryScriptPrompt({
+                  nodeTitle: normalizeText(parsed.data.nodeTitle),
+                  prompt,
+                  imageCount: imageUrls.length,
+                  videoCount: videoUrls.length,
+                })
+              : buildAgentPlanPrompt({
+                  nodeTitle: normalizeText(parsed.data.nodeTitle),
+                  prompt,
+                  imageCount: imageUrls.length,
+                  videoCount: videoUrls.length,
+                });

    const inlinedImageReferences = useVision
      ? (
@ -554,6 +563,8 @@ export async function executeWorkflowAnalysisResponse(body: unknown, signal?: Ab
      ? [...inlinedImageReferences, ...extractedVideoFrames].slice(0, 6)
      : [];

+    if (parsed.data.mode === "director-reconstruct" && !referenceImages.length)
+      throw Error("识图重建需要一张可读取的图片");
    const { text, finishReason } = await generateText({
      model,
      messages: [
@ -566,12 +577,11 @@ export async function executeWorkflowAnalysisResponse(body: unknown, signal?: Ab
        },
      ],
      temperature: 0.2,
-      maxOutputTokens:
-        parsed.data.mode === "story-script"
-          ? 12000
-          : parsed.data.mode === "video-analysis"
-            ? 2200
-            : 1600,
+      maxOutputTokens: ["story-script", "director-reconstruct"].includes(parsed.data.mode)
+        ? 12000
+        : parsed.data.mode === "video-analysis"
+          ? 2200
+          : 1600,
      maxRetries: 0,
      abortSignal: signal
        ? AbortSignal.any([signal, AbortSignal.timeout(150_000)])
@ -593,12 +603,25 @@ export async function executeWorkflowAnalysisResponse(body: unknown, signal?: Ab
    }
    const summary = normalizeText(parsedJson.summary) || normalizeText(text).slice(0, 160);

+    if (parsed.data.mode === "director-reconstruct")
+      return NextResponse.json({
+        success: true,
+        data: {
+          ...parseReconstructedScene(parsedJson),
+          raw: text,
+          model: selectedModel.model.displayName,
+        },
+      });
    if (parsed.data.mode === "video-analysis") {
      const analysisRows = normalizeVideoAnalysisRows(
        parsedJson.analysisRows,
        extractedVideoFrames,
      );
-      if (!analysisRows.length) return NextResponse.json({ success: false, error: "视频解析没有返回有效镜头，请重新解析" }, { status: 502 });
+      if (!analysisRows.length)
+        return NextResponse.json(
+          { success: false, error: "视频解析没有返回有效镜头，请重新解析" },
+          { status: 502 },
+        );
      const shotIdeas =
        normalizeShotIdeas(parsedJson.shotIdeas).length > 0
          ? normalizeShotIdeas(parsedJson.shotIdeas)
--- a/scripts/configure-aliyun-canvas-audio.cjs
+++ b/scripts/configure-aliyun-canvas-audio.cjs
@ -0,0 +1,91 @@
+// Run on the application host: APP_ROOT=/var/www/ai-chatbot node SCRIPT [--apply] [--music-enabled]
+// Only model rows are backed up; credentials never leave the existing provider.
+const fs = require("node:fs");
+const root = process.env.APP_ROOT || process.cwd();
+require(root + "/node_modules/dotenv").config({ path: root + "/.env.local", quiet: true });
+const sql = require(root + "/node_modules/postgres")(process.env.POSTGRES_URL, { max: 1 });
+const providerId = "7bce5924-bfd2-4e31-9f1d-825bd9d9f8b1";
+const apply = process.argv.includes("--apply"),
+  musicEnabled = process.argv.includes("--music-enabled");
+const models = [
+  {
+    modelId: "qwen3-tts-instruct-flash",
+    displayName: "千问3 TTS 指令版（阿里云官方）",
+    kind: "tts",
+    enabled: true,
+    defaults: { voice: "Cherry", language: "Chinese", format: "wav" },
+  },
+  {
+    modelId: "qwen3-tts-flash",
+    displayName: "千问3 TTS（阿里云官方）",
+    kind: "tts",
+    enabled: true,
+    defaults: { voice: "Cherry", language: "Chinese", format: "wav" },
+  },
+  {
+    modelId: "fun-music-v1",
+    displayName: "Fun-Music（阿里云官方·需邀测）",
+    kind: "music",
+    enabled: musicEnabled,
+    defaults: { instrumental: true, format: "mp3" },
+  },
+];
+(async () => {
+  const [provider] =
+    await sql`select id,type,"baseUrl","isEnabled" from "AiProvider" where id=${providerId}`;
+  if (
+    !provider ||
+    provider.type !== "alibaba" ||
+    !provider.isEnabled ||
+    new URL(provider.baseUrl).origin !== "https://dashscope.aliyuncs.com"
+  )
+    throw Error("Expected enabled official Alibaba Beijing provider");
+  await sql.begin(async (tx) => {
+    await tx`select id from "AiProvider" where id=${providerId} for update`;
+    const before =
+      await tx`select * from "AiModel" where "providerId"=${providerId} and "modelId" in ${tx(models.map((m) => m.modelId))}`;
+    if (apply) {
+      const dir = root + "/.model-config-backups";
+      fs.mkdirSync(dir, { recursive: true, mode: 0o700 });
+      fs.writeFileSync(
+        dir + "/canvas-audio-" + Date.now() + ".json",
+        JSON.stringify({ providerId, models: before }, null, 2),
+        { mode: 0o600 },
+      );
+    }
+    for (const model of models) {
+      const existing = before.filter((row) => row.modelId === model.modelId);
+      if (existing.length > 1) throw Error("Duplicate model rows: " + model.modelId);
+      const row = existing[0];
+      // Preserve an already enabled music row on subsequent idempotent runs.
+      const enabled = model.enabled || !!row?.isEnabled;
+      const capabilities = {
+        ...(row?.capabilities || {}),
+        kind: "audio",
+        audioType: model.kind,
+        official: true,
+        activationRequired: model.kind === "music" && !enabled,
+      };
+      const defaults = { ...(row?.defaultParams || {}), ...model.defaults };
+      if (apply) {
+        if (row)
+          await tx`update "AiModel" set "displayName"=${model.displayName},"isEnabled"=${enabled},"isHidden"=true,capabilities=${tx.json(capabilities)},"defaultParams"=${tx.json(defaults)},"updatedAt"=now() where id=${row.id}`;
+        else
+          await tx`insert into "AiModel" ("providerId","modelId","displayName","isEnabled","isHidden",capabilities,"defaultParams","sortOrder") values (${providerId},${model.modelId},${model.displayName},${enabled},true,${tx.json(capabilities)},${tx.json(defaults)},${model.kind === "music" ? 902 : model.modelId.includes("instruct") ? 900 : 901})`;
+      }
+      console.log(
+        JSON.stringify({
+          mode: apply ? "applied" : "dry-run",
+          modelId: model.modelId,
+          enabled,
+          activationRequired: capabilities.activationRequired,
+        }),
+      );
+    }
+  });
+})()
+  .catch((e) => {
+    console.error(e.message?.replace(/postgres:\/\/[^ ]+/g, "[redacted]") || e.name);
+    process.exitCode = 1;
+  })
+  .finally(() => sql.end());
--- a/scripts/qa/video-canvas-audio.cjs
+++ b/scripts/qa/video-canvas-audio.cjs
@ -0,0 +1,126 @@
+// Exercises the actual audio adapter with local provider/storage doubles; no paid calls.
+const fs = require("node:fs"),
+  path = require("node:path"),
+  assert = require("node:assert/strict");
+(async () => {
+  const esbuild = require(
+    path.resolve(
+      "node_modules/.pnpm",
+      fs.readdirSync("node_modules/.pnpm").find((n) => n.startsWith("esbuild@0.25.")),
+      "node_modules/esbuild",
+    ),
+  );
+  fs.mkdirSync(".qa/audio-expansion", { recursive: true });
+  const modules = {
+    "server-only": "",
+    "@/lib/db/queries":
+      "export async function getEnabledAiModelsWithProvider(){return globalThis.__audioModels}",
+    "@/lib/storage/service":
+      "export async function uploadFile(name,bytes){return globalThis.__audioUpload(name,bytes)}",
+    "./download-media":
+      "export async function downloadCanvasMedia(url,file){return globalThis.__audioDownload(url,file)}",
+  };
+  await esbuild.build({
+    entryPoints: ["lib/video-canvas/audio-generation.ts"],
+    outfile: ".qa/audio-expansion/adapter.cjs",
+    bundle: true,
+    platform: "node",
+    format: "cjs",
+    packages: "external",
+    alias: { "@": process.cwd() },
+    plugins: [
+      {
+        name: "qa",
+        setup(b) {
+          b.onResolve({ filter: /.*/ }, (args) =>
+            Object.hasOwn(modules, args.path) ? { path: args.path, namespace: "qa" } : undefined,
+          );
+          b.onLoad({ filter: /.*/, namespace: "qa" }, (args) => ({ contents: modules[args.path] }));
+        },
+      },
+    ],
+  });
+  const { executeCanvasAudio } = require(path.resolve(".qa/audio-expansion/adapter.cjs"));
+  global.__audioModels = [
+    {
+      model: { id: "official", modelId: "qwen3-tts-instruct-flash" },
+      provider: {
+        type: "alibaba",
+        baseUrl: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+        apiKey: "local-fixture",
+      },
+    },
+  ];
+  let posts = 0,
+    downloads = 0,
+    uploads = 0,
+    checkpoint;
+  global.fetch = async (url, init) => {
+    posts++;
+    assert.equal(init.method, "POST");
+    assert.equal(JSON.parse(init.body).input.voice, "Cherry");
+    return Response.json({
+      request_id: "fixture",
+      output: {
+        audio: {
+          url: "https://example.com/test.wav",
+          expires_at: Math.floor(Date.now() / 1000) + 3600,
+        },
+      },
+    });
+  };
+  global.__audioDownload = async (url, file) => {
+    downloads++;
+    assert.ok(checkpoint);
+    fs.writeFileSync(file, Buffer.from("RIFF1234WAVEfixtureaudio"));
+  };
+  global.__audioUpload = async (name, bytes) => {
+    uploads++;
+    if (uploads === 1) throw Error("storage unavailable");
+    assert.ok(bytes.byteLength);
+    return { url: "https://storage.example.com/archived.wav" };
+  };
+  const input = { mode: "tts", modelRef: "official", text: "测试", voice: "Cherry" };
+  const lifecycle = {
+    assertActive: async () => {},
+    checkpoint: async (value) => {
+      checkpoint = value.upstreamAudio;
+    },
+  };
+  await assert.rejects(() => executeCanvasAudio(input, lifecycle), /转存暂未完成/);
+  assert.equal(posts, 1);
+  assert.ok(checkpoint);
+  const result = await executeCanvasAudio(input, { ...lifecycle, resume: checkpoint });
+  assert.equal(result.url, "https://storage.example.com/archived.wav");
+  assert.equal(posts, 1);
+  assert.equal(downloads, 2);
+  await assert.rejects(
+    () => executeCanvasAudio(input, { ...lifecycle, resume: { ...checkpoint, expiresAt: 1 } }),
+    /过期/,
+  );
+  assert.equal(posts, 1);
+  global.fetch = async () => {
+    posts++;
+    throw Error("connection lost");
+  };
+  await assert.rejects(() => executeCanvasAudio(input, lifecycle), /不会自动重发/);
+  assert.equal(posts, 2);
+  global.fetch = async () => Response.json({ code: "Forbidden" }, { status: 403 });
+  await assert.rejects(() => executeCanvasAudio(input, lifecycle), /403/);
+  await assert.rejects(
+    () =>
+      executeCanvasAudio(input, {
+        ...lifecycle,
+        assertActive: async () => {
+          throw Error("lost lease");
+        },
+      }),
+    /lost lease/,
+  );
+  console.log(
+    "PASS: actual adapter body, checkpoint-before-download, archival recovery without POST, expiry, transport uncertainty, provider failure and lease fencing",
+  );
+})().catch((e) => {
+  console.error(e);
+  process.exitCode = 1;
+});
--- a/scripts/qa/video-canvas-task-db.cjs
+++ b/scripts/qa/video-canvas-task-db.cjs
@ -295,6 +295,7 @@ const fs = require("fs"),
    fs.writeFileSync(callsFile, "");
    const unexpected = `throw Error("unexpected provider path")`;
    const stubs = {
+      "./audio-generation": `export const executeCanvasAudio=async()=>{${unexpected}};`,
      "@/lib/ai/tools/generate-video": `export const executeDirectVideoGeneration=async()=>{${unexpected}};`,
      "@/lib/canvas/generate-image-response": `export const executeCanvasImageResponse=async()=>{${unexpected}};`,
      "./workflow-analysis-response": `export const executeWorkflowAnalysisResponse=async()=>{${unexpected}};`,
--- a/tests/routes/video-canvas-director.test.ts
+++ b/tests/routes/video-canvas-director.test.ts
@ -189,6 +189,8 @@ test("recording stops tracks and rejects partial exports on cancel or background
    "cancel",
    "hidden",
    "render-error",
+    "warmup-cancel",
+    "warmup-hidden",
  ] as const) {
    const saved = new Map<string, PropertyDescriptor | undefined>();
    const install = (key: string, value: unknown) => {
@ -251,17 +253,25 @@ test("recording stops tracks and rejects partial exports on cancel or background
        callbacks.clear();
        pending.forEach((fn) => fn(time));
      };
+      const warming = outcome.startsWith("warmup-");
+      if (!warming) {
+        step(10);
+        step(20);
+        await Promise.resolve();
+      }
      if (outcome === "complete" || outcome === "complete-webm") {
        step(1000);
        step(3000);
+        step(3033);
+        step(3066);
        expect((await result).type).toBe(
          outcome === "complete-webm" ? "video/webm;codecs=vp8" : "video/mp4",
        );
        expect(rendered.at(-1)).toBe(2);
      } else {
        const rejected = expect(result).rejects.toThrow();
-        if (outcome === "cancel") controller.abort();
-        else if (outcome === "hidden") {
+        if (outcome === "cancel" || outcome === "warmup-cancel") controller.abort();
+        else if (outcome === "hidden" || outcome === "warmup-hidden") {
          doc.hidden = true;
          doc.dispatchEvent(new Event("visibilitychange"));
        } else {
@ -270,7 +280,7 @@ test("recording stops tracks and rejects partial exports on cancel or background
        }
        await rejected;
      }
-      expect(stopped).toBe(1);
+      expect(stopped).toBe(warming ? 0 : 1);
      expect(callbacks.size).toBe(0);
    } finally {
      for (const [key, descriptor] of saved) {
@ -403,3 +413,116 @@ test("whole-path transforms change camera and target together while preserving d
    motion.keyframes[0].target[2] + 3,
  ]);
 });
+
+import {
+  recordObjectFrames,
+  sampleDirectorObjects,
+  readObjectTracks,
+} from "@/lib/video-canvas/director-object-tracks";
+import {
+  parseReconstructedScene,
+  insertReconstructedScene,
+} from "@/lib/video-canvas/director-reconstruction";
+test("object tracks interpolate independently, persist, mute and preserve shortest rotation", () => {
+  let scene = defaultDirectorScene();
+  const id = scene.objects[0].id;
+  scene = recordObjectFrames(scene, [id], 0, () => "key-0");
+  scene.objects = scene.objects.map((o) =>
+    o.id === id
+      ? {
+          ...o,
+          position: [10, 2, 4],
+          rotation: [0, Math.PI * 2 - 0.2, 0],
+          scale: [2, 3, 4],
+          visible: false,
+        }
+      : o,
+  );
+  scene = recordObjectFrames(scene, [id], scene.motion.duration, () => "key-1");
+  const mid = sampleDirectorObjects(scene, scene.motion.duration / 2)[0];
+  expect(mid.position[0]).toBeCloseTo((scene.objectTracks![id].keyframes[0].position[0] + 10) / 2);
+  expect(mid.rotation[1]).toBeCloseTo(-0.1);
+  expect(mid.visible).toBe(true);
+  expect(sampleDirectorObjects(scene, scene.motion.duration)[0].visible).toBe(false);
+  const restored = readDirectorScene(JSON.parse(JSON.stringify(scene)));
+  expect(restored.objectTracks).toEqual(scene.objectTracks);
+  scene.objectTracks![id].muted = true;
+  expect(sampleDirectorObjects(scene, 0)[0]).toEqual(scene.objects[0]);
+});
+test("malformed/orphan object tracks are excluded and recording replaces exact frame", () => {
+  let scene = defaultDirectorScene();
+  const id = scene.objects[0].id;
+  scene = recordObjectFrames(scene, [id], 0, () => "one");
+  scene = recordObjectFrames(scene, [id], 0, () => "two");
+  expect(scene.objectTracks![id].keyframes).toHaveLength(1);
+  const valid = scene.objectTracks![id];
+  expect(
+    readObjectTracks(
+      {
+        ghost: valid,
+        [id]: {
+          ...valid,
+          keyframes: [...valid.keyframes, { ...valid.keyframes[0], id: "bad", time: Infinity }],
+        },
+      },
+      scene.objects,
+      scene.motion.duration,
+    ),
+  ).toEqual({ [id]: valid });
+  expect(() => recordObjectFrames(scene, [id], -1, () => "bad")).toThrow();
+});
+test("reference reconstruction validates geometry and merges without clobbering original objects", () => {
+  const generated = parseReconstructedScene({
+    summary: "Approximate geometry",
+    objects: [
+      {
+        name: "table",
+        kind: "box",
+        position: [0, 0.5, 0],
+        rotation: [0, 0, 0],
+        scale: [2, 1, 1],
+        color: "#ffffff",
+      },
+    ],
+    camera: { position: [3, 3, 3], target: [0, 0.5, 0], fov: 45 },
+  });
+  const scene = defaultDirectorScene();
+  let count = 0;
+  const insert = insertReconstructedScene(
+    scene,
+    generated.directorScene,
+    false,
+    () => `added-${count++}`,
+  );
+  expect(insert.objects.length).toBe(scene.objects.length + 1);
+  expect(insert.objects[0]).toEqual(scene.objects[0]);
+  expect(new Set(insert.objects.map((o) => o.id)).size).toBe(insert.objects.length);
+  const replaced = insertReconstructedScene(
+    scene,
+    generated.directorScene,
+    true,
+    () => `new-${count++}`,
+  );
+  expect(replaced.objects).toHaveLength(1);
+  expect(replaced.objectTracks).toEqual({});
+  expect(() =>
+    parseReconstructedScene({
+      summary: "bad",
+      objects: [],
+      camera: { position: [0, 0, 0], target: [0, 0, 0], fov: 45 },
+    }),
+  ).toThrow();
+});
+
+test("global duration rescales object tracks and deleting an object deletes its track", () => {
+  let scene = defaultDirectorScene();
+  const id = scene.objects[0].id;
+  scene = recordObjectFrames(scene, [id], scene.motion.duration, () => "last");
+  const old = scene.motion.duration;
+  const changed = commitDirectorScene(scene, {
+    ...scene,
+    motion: { ...scene.motion, duration: old * 2 },
+  });
+  expect(changed.objectTracks![id].keyframes[0].time).toBe(old * 2);
+  expect(commitDirectorScene(changed, { ...changed, objects: [] }).objectTracks).toEqual({});
+});
--- a/tests/routes/video-canvas-tasks.test.ts
+++ b/tests/routes/video-canvas-tasks.test.ts
@ -656,3 +656,94 @@ test("video recovery requires an original server checkpoint and matching selecte
  }
  expect(canResumeCanvasTask({ ...task, attempts: 20 }, recoveryTime)).toBe(false);
 });
+
+import { buildAudioRequest, officialAudioBase } from "@/lib/video-canvas/audio-contract";
+import { getAudioCheckpoint } from "@/lib/video-canvas/task-recovery";
+test("official Alibaba audio requests match each model capabilities", () => {
+  const tts = buildAudioRequest(
+    { mode: "tts", modelRef: "a", text: "你好", voice: "Cherry", instructions: "温柔自然" },
+    "qwen3-tts-instruct-flash",
+  );
+  expect(tts.path).toContain("multimodal-generation");
+  expect(tts.body.input).toMatchObject({
+    text: "你好",
+    voice: "Cherry",
+    instructions: "温柔自然",
+    optimize_instructions: true,
+  });
+  expect(tts.body.input).not.toHaveProperty("rate");
+  expect(() =>
+    buildAudioRequest(
+      { mode: "tts", modelRef: "a", text: "你好", voice: "Cherry", rate: 2 },
+      "qwen3-tts-flash",
+    ),
+  ).toThrow("不支持数值");
+  expect(() =>
+    buildAudioRequest(
+      { mode: "tts", modelRef: "a", text: "你好", voice: "unsupported" },
+      "qwen3-tts-flash",
+    ),
+  ).toThrow("音色");
+  const modern = buildAudioRequest(
+    { mode: "tts", modelRef: "a", text: "你好", voice: "longanlingxin", format: "mp3", rate: 1.5 },
+    "qwen-audio-3.0-tts-plus",
+  );
+  expect(modern.body.input).toMatchObject({ format: "mp3", rate: 1.5, language_hints: ["zh"] });
+});
+test("music validates lyrics and excludes singing fields from instrumental output", () => {
+  const music = buildAudioRequest(
+    {
+      mode: "music",
+      modelRef: "a",
+      text: "轻快吉他",
+      lyrics: "不应该传的歌词",
+      instrumental: true,
+      format: "mp3",
+    },
+    "fun-music-v1",
+  );
+  expect(music.body.input).toEqual({ prompt: "轻快吉他", is_instrumental: true, format: "mp3" });
+  expect(() =>
+    buildAudioRequest(
+      { mode: "music", modelRef: "a", lyrics: "太短", instrumental: false },
+      "fun-music-v1",
+    ),
+  ).toThrow("歌词");
+  expect(() =>
+    buildAudioRequest({ mode: "music", modelRef: "a", text: "音乐" }, "qwen3-tts-flash"),
+  ).toThrow("不匹配");
+});
+test("audio credentials can only target official domains and correct region", () => {
+  expect(
+    officialAudioBase("https://dashscope.aliyuncs.com/compatible-mode/v1", "fun-music-v1"),
+  ).toBe("https://dashscope.aliyuncs.com");
+  expect(() => officialAudioBase("https://dashscope.aliyuncs.com.evil.test")).toThrow();
+  expect(() => officialAudioBase("http://dashscope.aliyuncs.com")).toThrow();
+  expect(() => officialAudioBase("https://dashscope-intl.aliyuncs.com", "fun-music-v1")).toThrow(
+    "北京",
+  );
+});
+test("only matching persisted audio result may resume archival", () => {
+  const task = {
+    request: { kind: "audio", params: { modelRef: "ref" } },
+    result: {
+      upstreamAudio: {
+        operation: "audio",
+        modelRef: "ref",
+        modelId: "qwen3-tts-flash",
+        url: "https://example.com/audio.wav",
+        format: "wav",
+      },
+    },
+  };
+  expect(getAudioCheckpoint(task)).toMatchObject({ modelRef: "ref" });
+  expect(
+    getAudioCheckpoint({ ...task, request: { kind: "audio", params: { modelRef: "another" } } }),
+  ).toBeNull();
+  expect(
+    getAudioCheckpoint({
+      ...task,
+      result: { upstreamAudio: { ...task.result.upstreamAudio, url: "file:///tmp/private" } },
+    }),
+  ).toBeNull();
+});