fix: surface resilient subject asset generation

2026-05-18 18:15:45 +08:00
parent cc4c021074
commit 095c6f1c00
8 changed files with 251 additions and 177 deletions
--- a/web/components/ad-recreation-board.tsx
+++ b/web/components/ad-recreation-board.tsx
@@ -38,6 +38,7 @@ import {
  generateProductAngleAsset,
  generateSubjectAssets,
  generatedImageUrl,
+  getJob,
  getRuntimeHealth,
  hasCutout,
  listCharacterLibrary,
@@ -570,9 +571,9 @@ function audioModelTrace(models?: RuntimeModels): ModelTraceSpec {
    title: "音频解析",
    model: modelList([models?.asr, models?.translate, models?.asr_fallback]),
    chain: [
-      `ASR 转写：优先 ${modelValue(models?.asr)}，失败后尝试本机 ${modelValue(models?.local_asr)}，再回退 ${modelValue(models?.asr_fallback)}`,
-      `字幕翻译：${modelValue(models?.translate)} 输出中文逐句时间轴`,
-      `讲话人 / 节奏 / 背景音：${modelValue(models?.asr_fallback)} 读取 audio.wav 做多模态音频分析`,
+      `ASR 转写：优先 ${modelValue(models?.asr)}；失败后尝试本机 ${modelValue(models?.local_asr)}；仍失败才回退 ${modelValue(models?.asr_fallback)}，并拒绝假字幕/重复时间轴`,
+      `字幕翻译：${modelValue(models?.translate)} 按 ASR 段落输出中文；失败时保留原文时间轴，中文可为空`,
+      `讲话人 / 节奏 / 背景音：${modelValue(models?.asr_fallback)} 读取 audio.wav + 转写时间轴做多模态分析；失败时用本地时长/段落估算兜底`,
    ],
    note: "点击“解析音频”后触发；开始任务下载完成后也会自动走这条链路。",
  }
@@ -583,9 +584,10 @@ function productModelTrace(models?: RuntimeModels): ModelTraceSpec {
    title: "产品视角识别 / 补图",
    model: modelList([models?.product_view, models?.image]),
    chain: [
-      `批量视角识别：${modelValue(models?.product_view)} 一次读取同一产品多张图，标注视角、左右、上下、用途和风险`,
-      `缺角度补图：${imageModelChain(models)} 读取最相关的多张已上传参考图，按同一肩颈按摩仪结构补齐缺失视角`,
-      "前端只保存标注和 AI 补图结果；后续生成视频时每条最多挑 6 张相关产品图",
+      `批量视角识别：${modelValue(models?.product_view)} 多图读取同一产品素材，标注视角、佩戴者左右、上下、内外侧、用途和风险`,
+      "识别兜底：批量失败会按单图重试；仍失败或文件缺失时写入本地默认视角，并在 risk/note 标明兜底原因",
+      `缺角度补图：${imageModelChain(models)} 走 /images/edits，最多读取 6 张已上传参考图补齐缺失视角；失败保留重试入口，不自动换模型`,
+      "前端只保存标注和 AI 补图结果；后续首尾帧/视频规划每条最多挑 6 张相关产品图",
    ],
    note: "上传产品图、重新识别、缺视角重试都会使用这组模型链路。",
  }
@@ -594,11 +596,11 @@ function productModelTrace(models?: RuntimeModels): ModelTraceSpec {
 function similarSubjectModelTrace(models: RuntimeModels | undefined, subjectStyle: SubjectStyleMode): ModelTraceSpec {
  return {
    title: subjectStyle === "transparent_human" ? "相似透明骨架主体" : "相似普通真人主体",
-    model: subjectImageModelChain(models),
+    model: modelList([models?.vision, models?.subject_image]),
    chain: [
-      "参考策略：先用视觉模型把关键帧/模板转成非身份化文字 brief，生图请求不再上传参考图",
+      `视觉 brief：${modelValue(models?.vision)} 把关键帧/模板图转成非身份化文字 brief；失败时继续用用户方向和模板文字`,
      `主体类型：${subjectStyle === "transparent_human" ? "透明/半透明皮肤包裹可见白色骨架" : "普通商业广告真人"}`,
-      `图像生成：${subjectImageModelChain(models)} 走 /images/generations 逐张生成高清图，视图数量由“全部/常用/自定义”决定`,
+      `图像生成：${subjectImageModelChain(models)} 走 /images/generations 逐张文字生图；当前 similar 模式不上传原帧或模板图作为 image-edit 参考`,
      "身份锁定：整套图必须是同一个主体，性别表现、年龄段、体型、材质和风格保持一致",
    ],
    note: "这是生成类似但创新的主体，不是复制、抠出或复刻源视频人物身份；内置形象也只作为方向参考。",
@@ -611,8 +613,8 @@ function scriptRewriteModelTrace(models?: RuntimeModels): ModelTraceSpec {
    model: modelList([models?.audio_rewrite, models?.asr_fallback, models?.translate]),
    chain: [
      `主改写：${modelValue(models?.audio_rewrite)} 根据原文案、当前分镜、作者想法生成新口播`,
-      `失败回退：依次尝试 ${modelValue(models?.asr_fallback)} 和 ${modelValue(models?.translate)}`,
-      "返回结果只写入当前分镜文案编辑框；生成视频时再把当前文案写入分镜 action",
+      `模型回退：依次尝试 ${modelValue(models?.asr_fallback)} 和 ${modelValue(models?.translate)}；全部失败时用本地模板保留分镜可编辑`,
+      "返回结果只写入当前分镜文案编辑框；点击保存规划后才写入 frame.storyboard.action",
    ],
  }
 }
@@ -625,8 +627,9 @@ function videoModelTrace(models: RuntimeModels | undefined, model: string): Mode
      `前端选择：${model}`,
      `后端解析：${resolveVideoModelLabel(models, model)}`,
      `服务商：${modelValue(models?.video_provider)} · ${modelValue(models?.video_base_url)}`,
-      "输入：已确认的首尾帧、当前分镜文案、产品素材、相似主体资产和画面规划",
-      "输出：异步候选视频，完成后回填到对应分镜行",
+      "当前主工作台暂停直接提交视频；旧入口误触也会被页面层保护",
+      "开放后输入会包含已确认首尾帧、当前分镜文案、产品素材、相似主体资产和画面规划",
+      "输出为异步候选视频，完成后回填到对应分镜行；Sora 已停用",
    ],
  }
 }
@@ -2144,7 +2147,7 @@ function SourceReferenceBuildPanel({
  onJobUpdate: (job: Job) => void
  runtimeModels?: RuntimeModels
 }) {
-  const [subjectBusy, setSubjectBusy] = useState(false)
+  const [subjectBusyFor, setSubjectBusyFor] = useState<{ jobId: string; jobLabel: string; viewCount: number } | null>(null)
  const [subjectAssetBusy, setSubjectAssetBusy] = useState<string | null>(null)
  const [subjectMode, setSubjectMode] = useState<SubjectMode>("source_similar")
  const [subjectStyle, setSubjectStyle] = useState<SubjectStyleMode>("transparent_human")
@@ -2220,6 +2223,7 @@ function SourceReferenceBuildPanel({
    ? `${selectedCharacter.name} · 模板参考`
    : "源视频关键帧 · 相似创新"
  const templateRequired = subjectMode === "template" && !selectedSubjectTemplate && !selectedCharacter
+  const subjectBusy = !!subjectBusyFor
  const generationCtaLabel = subjectMode === "template"
    ? `用模板生成 ${selectedSubjectViews.length} 张主体视图`
    : `从源视频创新生成 ${selectedSubjectViews.length} 张主体视图`
@@ -2265,13 +2269,14 @@ function SourceReferenceBuildPanel({
    }
    const baseFrame = subjectReferenceFrames[0]
    if (!baseFrame) return
-    setSubjectBusy(true)
+    const requestJobId = job.id
+    setSubjectBusyFor({ jobId: requestJobId, jobLabel: shortId(requestJobId), viewCount: selectedSubjectViews.length })
    try {
      let workingJob = job
      let workingFrame = workingJob.frames.find((frame) => frame.index === baseFrame.index) ?? baseFrame
      let element = workingFrame.elements?.find(isSimilarActorElement)
      if (!element) {
-        workingJob = await addElement(job.id, baseFrame.index, {
+        workingJob = await addElement(requestJobId, baseFrame.index, {
          name_zh: selectedTemplatePrompt
            ? `相似透明骨架主体 · ${selectedTemplatePrompt.name}`
            : subjectStyle === "transparent_human" ? "相似透明骨架主体" : "相似广告主角",
@@ -2288,7 +2293,7 @@ function SourceReferenceBuildPanel({
      }
      if (!element) throw new Error("similar subject element missing")

-      const updated = await generateSubjectAssets(job.id, baseFrame.index, element.id, {
+      const updated = await generateSubjectAssets(requestJobId, baseFrame.index, element.id, {
        subject_kind: "living",
        subject_style: subjectStyle,
        reconstruction_mode: "similar",
@@ -2304,9 +2309,12 @@ function SourceReferenceBuildPanel({
      onJobUpdate(updated)
      toast.success(`相似主体 ${selectedSubjectViews.length} 张高清白底图已生成`)
    } catch (e) {
+      try {
+        onJobUpdate(await getJob(requestJobId))
+      } catch { /* keep original error visible */ }
      toast.error("相似主体重构失败：" + (e instanceof Error ? e.message : String(e)))
    } finally {
-      setSubjectBusy(false)
+      setSubjectBusyFor(null)
    }
  }

@@ -2438,7 +2446,8 @@ function SourceReferenceBuildPanel({
            ))}
          </div>

-          <div className={`transition ${subjectMode === "source_similar" ? "pointer-events-none opacity-38 grayscale" : ""}`}>
+          {subjectMode === "template" ? (
+          <div>
            <div className="grid grid-cols-[repeat(auto-fill,minmax(120px,1fr))] gap-2">
              {subjectTemplateLibrary.map((template) => {
                const preview = characterPreviewImage(template)
@@ -2494,6 +2503,11 @@ function SourceReferenceBuildPanel({
              })}
            </div>
          </div>
+          ) : (
+            <div className="rounded-md border border-cyan-200/18 bg-cyan-300/[0.06] px-2.5 py-2 text-[10px] leading-snug text-cyan-50/62">
+              当前跳过模板库：本次只用源视频关键帧的文字化主体特征生成创新主体。模板卡已收起，避免占用生成结果区域。
+            </div>
+          )}

          {subjectMode === "template" && (selectedSubjectTemplate?.images?.length || selectedCharacter?.images?.length) ? (
            <div className="mt-2 flex gap-1.5 overflow-x-auto pb-0.5">
@@ -2545,6 +2559,45 @@ function SourceReferenceBuildPanel({
            </div>
            <ModelTrace trace={similarSubjectModelTrace(runtimeModels, subjectStyle)} compact />
          </div>
+          {subjectBusyFor ? (
+            <div className="mb-2 rounded-md border border-cyan-200/20 bg-cyan-300/[0.07] px-2.5 py-2 text-[10px] leading-snug text-cyan-50/70">
+              正在为素材 {subjectBusyFor.jobLabel} 生成 {subjectBusyFor.viewCount} 张主体视图；本次请求已锁定素材、参考帧、模式和视图数量，切换其他模块不会改变生成目标，完成后会回写到该素材。
+            </div>
+          ) : null}
+
+          {visibleActorAssets.length ? (
+            <div className="mb-2 grid grid-cols-[repeat(auto-fill,minmax(96px,1fr))] gap-2">
+              {visibleActorAssets.map((asset) => {
+                const busyMode = subjectAssetBusy?.endsWith(asset.id) ? subjectAssetBusy.split(":")[0] : ""
+                return (
+                  <MediaAssetTile
+                    key={asset.id}
+                    src={subjectAssetUrl(job, asset)}
+                    href={subjectAssetUrl(job, asset)}
+                    alt={asset.label || asset.view}
+                    label={asset.label || asset.view || "主体视图预览"}
+                    meta={asset.width && asset.height ? `${asset.width}x${asset.height}` : undefined}
+                    className="aspect-[9/16] w-20 bg-white 2xl:w-24"
+                    objectFit="contain"
+                    title={asset.label || asset.view}
+                    actions={[{
+                      key: "regen",
+                      label: "重新生成这一张",
+                      icon: <RefreshCw className="h-3 w-3" />,
+                      tone: "cyan",
+                      busy: busyMode === "regen",
+                      disabled: !!subjectAssetBusy,
+                      onClick: () => void regenerateSubjectAsset(asset),
+                    }]}
+                    onDelete={() => void deleteActorAsset(asset)}
+                    deleting={busyMode === "delete"}
+                    deleteDisabled={!!subjectAssetBusy}
+                    deleteLabel="删除这一张"
+                  />
+                )
+              })}
+            </div>
+          ) : null}

          <div className="grid gap-2 xl:grid-cols-[auto_auto_minmax(220px,1fr)_auto] xl:items-start">
            <div className="flex rounded-md border border-white/10 bg-black/28 p-0.5">
@@ -2595,7 +2648,7 @@ function SourceReferenceBuildPanel({
              className="inline-flex h-9 min-w-[170px] items-center justify-center gap-1 rounded-md bg-white px-3 text-[11px] font-semibold text-black transition hover:bg-white/90 disabled:cursor-not-allowed disabled:opacity-40"
            >
              {subjectBusy ? <Loader2 className="h-3.5 w-3.5 animate-spin" /> : <Sparkles className="h-3.5 w-3.5" />}
-              {generationCtaLabel}
+              {subjectBusyFor ? `生成中 · ${subjectBusyFor.jobLabel}` : generationCtaLabel}
            </button>
          </div>

@@ -2623,45 +2676,13 @@ function SourceReferenceBuildPanel({
            </div>
          ) : null}

-          {visibleActorAssets.length ? (
-            <div className="mt-2 grid grid-cols-[repeat(auto-fill,minmax(96px,1fr))] gap-2">
-              {visibleActorAssets.map((asset) => {
-                const busyMode = subjectAssetBusy?.endsWith(asset.id) ? subjectAssetBusy.split(":")[0] : ""
-                return (
-                  <MediaAssetTile
-                    key={asset.id}
-                    src={subjectAssetUrl(job, asset)}
-                    href={subjectAssetUrl(job, asset)}
-                    alt={asset.label || asset.view}
-                    label={asset.label || asset.view || "主体视图预览"}
-                    meta={asset.width && asset.height ? `${asset.width}x${asset.height}` : undefined}
-                    className="aspect-[9/16] w-20 bg-white 2xl:w-24"
-                    objectFit="contain"
-                    title={asset.label || asset.view}
-                    actions={[{
-                      key: "regen",
-                      label: "重新生成这一张",
-                      icon: <RefreshCw className="h-3 w-3" />,
-                      tone: "cyan",
-                      busy: busyMode === "regen",
-                      disabled: !!subjectAssetBusy,
-                      onClick: () => void regenerateSubjectAsset(asset),
-                    }]}
-                    onDelete={() => void deleteActorAsset(asset)}
-                    deleting={busyMode === "delete"}
-                    deleteDisabled={!!subjectAssetBusy}
-                    deleteLabel="删除这一张"
-                  />
-                )
-              })}
-            </div>
-          ) : (
+          {!visibleActorAssets.length ? (
            <div className="mt-2 rounded border border-dashed border-white/12 px-2 py-2 text-[10.5px] leading-snug text-white/32">
              {subjectMode === "template"
                ? "先选主体模板，再生成新主体视图；模板只作为文字化创意方向，不再作为强参考图复制。"
                : "直接使用关键帧的文字化主体特征生成创新主体；后端不会上传源图给生图端点。"}
            </div>
-          )}
+          ) : null}
        </div>
      </div>
    </div>
--- a/web/components/dashboard.tsx
+++ b/web/components/dashboard.tsx
@@ -673,7 +673,7 @@ export const Dashboard = forwardRef<DashboardHandle, Props>(function Dashboard({
            {key === "videogen" && (
              <>
                <KanbanCard tone="violet" tags={["SKG 网关"]} title="Seedance / Kling / Veo 3">
-                  <div className="text-[11px] text-[var(--text-soft)]">通过 /v1/videos 网关提交，模型 ID 走环境变量映射</div>
+                  <div className="text-[11px] text-[var(--text-soft)]">按后端 VIDEO_CREATE_PATHS 提交，模型 ID 走环境变量映射</div>
                </KanbanCard>
                <KanbanCard tone="violet" tags={["外部"]} title="Seedance">
                  <div className="text-[11px] text-[var(--text-soft)]">字节跳动 · 需独立 API key</div>