auto-save 2026-05-14 06:11 (~6)

.memory/worklog.json

@@ -1,19 +1,5 @@
 {
   "entries": [
-    {
-      "files_changed": 3,
-      "hash": "4779c26",
-      "message": "auto-save 2026-05-12 16:49 (~3)",
-      "ts": "2026-05-12T16:50:05+08:00",
-      "type": "commit"
-    },
-    {
-      "files_changed": 4,
-      "hash": "345391d",
-      "message": "auto-save 2026-05-12 16:55 (~4)",
-      "ts": "2026-05-12T16:55:37+08:00",
-      "type": "commit"
-    },
     {
       "files_changed": 3,
       "hash": "4138bea",
@@ -3354,6 +3340,19 @@
       "type": "session-heartbeat",
       "message": "Claude 会话活跃 · 最近命令：claude · 1 项未提交变更 · 最近提交：auto-save 2026-05-14 06:00 (~2)",
       "files_changed": 1
+    },
+    {
+      "ts": "2026-05-14T06:05:57+08:00",
+      "type": "commit",
+      "message": "auto-save 2026-05-14 06:05 (~1)",
+      "hash": "2b54616",
+      "files_changed": 1
+    },
+    {
+      "ts": "2026-05-13T22:08:51Z",
+      "type": "session-heartbeat",
+      "message": "Codex 会话活跃 · 最近命令：codex · 4 项未提交变更 · 最近提交：auto-save 2026-05-14 06:05 (~1)",
+      "files_changed": 4
     }
   ]
 }

api/main.py

@@ -97,6 +97,8 @@ AssetSize = Literal["source", "1024", "1536", "2048"]
 AssetQuality = Literal["hd"]
 SubjectKind = Literal["object", "living"]
 SubjectView = str
+SceneMode = Literal["remove_subject", "similar", "style"]
+SceneStyle = Literal["source", "premium_product", "clean_studio", "warm_lifestyle", "cinematic"]
 FRAME_TARGET_LABELS: dict[FrameExtractTarget, str] = {
     "balanced": "综合关键帧",
     "subject": "清晰主体",
@@ -191,6 +193,8 @@ class SceneAsset(BaseModel):
     height: int = 0
     quality: AssetQuality = "hd"
     size: AssetSize = "source"
+    scene_mode: SceneMode = "remove_subject"
+    scene_style: SceneStyle = "source"
     quality_report: QualityReport | None = None
     created_at: float = 0.0
 
@@ -1930,6 +1934,8 @@ class UpdateElementReq(BaseModel):
 class GenerateSceneAssetReq(BaseModel):
     quality: AssetQuality = "hd"
     size: AssetSize = "source"
+    scene_mode: SceneMode = "remove_subject"
+    scene_style: SceneStyle = "source"
 
 
 class GenerateSubjectAssetsReq(BaseModel):
@@ -2058,7 +2064,8 @@ def delete_element(job_id: str, idx: int, element_id: str) -> Job:
 
 @app.post("/jobs/{job_id}/frames/{idx}/scene-asset", response_model=Job)
 def generate_scene_asset(job_id: str, idx: int, req: GenerateSceneAssetReq) -> Job:
-    """为关键帧生成一张干净、高清的场景参考图。默认一帧只需要一张，重跑会保留历史供人工比对。"""
+    """为关键帧生成一张干净、高清的场景参考图。默认一帧只需要一张，重跑会保留历史供人工比对。
+    场景图排在主体资产之后：优先依据已确认主体，去主体并补全背景，再按模式生成原场景/相似场景/换风格场景。"""
     import time as _time
     job = JOBS.get(job_id)
     if not job:
@@ -2068,12 +2075,51 @@ def generate_scene_asset(job_id: str, idx: int, req: GenerateSceneAssetReq) -> J
     if not src.exists():
         raise HTTPException(404, "source frame file missing")
 
+    confirmed_subjects = [
+        (e.name_en or e.name_zh).strip()
+        for e in (frame.elements or [])
+        if (e.subject_assets or [])
+    ]
+    if not confirmed_subjects:
+        confirmed_subjects = [
+            (e.name_en or e.name_zh).strip()
+            for e in (frame.elements or [])
+            if (e.name_en or e.name_zh).strip()
+        ][:3]
+    subject_clause = (
+        "Confirmed foreground subject(s) to remove: " + ", ".join(confirmed_subjects) + ". "
+        if confirmed_subjects
+        else "Remove the main foreground subject from the frame if present. "
+    )
+    mode_clause = {
+        "remove_subject": (
+            "Keep the original environment, camera angle, perspective, composition, lighting direction, color mood, and spatial layout. "
+            "The result should be an empty clean scene/background plate with the subject removed and the occluded background reconstructed."
+        ),
+        "similar": (
+            "Create a similar but not identical scene/background plate: keep the same camera angle, rough spatial layout, lighting direction, and usage context, "
+            "but vary props, surface details, textures, and small environmental details so it is not a duplicate of the source."
+        ),
+        "style": (
+            "Create a scene/background plate with the same camera angle and spatial layout, but reinterpret the environment in the selected visual style. "
+            "Keep it believable and useful for image-to-video generation."
+        ),
+    }[req.scene_mode]
+    style_clause = {
+        "source": "Follow the original source style.",
+        "premium_product": "Use a premium product-advertising style: polished, high-end, clean commercial lighting, refined materials.",
+        "clean_studio": "Use a clean studio style: simple surfaces, controlled lighting, minimal distractions.",
+        "warm_lifestyle": "Use a warm lifestyle style: realistic lived-in details, soft natural light, approachable atmosphere.",
+        "cinematic": "Use a cinematic style: dramatic but natural lighting, richer depth, filmic contrast, not fantasy.",
+    }[req.scene_style]
     prompt = (
-        "Create one clean high-definition scene reference image from this frame. "
-        "Remove watermarks, platform UI, captions, usernames, hashtags, logos, and overlay graphics. "
-        "Preserve the original camera angle, composition, environment, lighting style, and believable spatial layout. "
-        "Do not create multiple views. Do not isolate objects. Keep it useful as the scene/background reference for image-to-video generation. "
-        "Enhance clarity and texture while avoiding over-smoothing or changing important visual details."
+        "Create one clean high-definition scene/background reference image from this frame. "
+        + subject_clause
+        + "Do not include the removed subject, duplicate people, animals, products, text, watermark, platform UI, captions, usernames, hashtags, logos, or overlay graphics. "
+        + mode_clause + " "
+        + style_clause + " "
+        + "Enhance clarity and texture while avoiding over-smoothing, warped geometry, or changing important perspective details. "
+        + "Do not create multiple views. Do not isolate objects."
     )
     models = [IMAGE_MODEL, "gemini-3.1-flash-image-preview", "gemini-2.5-flash-image"]
     try:
@@ -2093,6 +2139,8 @@ def generate_scene_asset(job_id: str, idx: int, req: GenerateSceneAssetReq) -> J
         height=height,
         quality=req.quality,
         size=req.size,
+        scene_mode=req.scene_mode,
+        scene_style=req.scene_style,
         quality_report=report,
         created_at=_time.time(),
     )

docs/source-analysis.html

@@ -555,7 +555,7 @@
           <div class="step"><div class="num">2</div><h3>镜头拆解</h3><p>拆轨、抽关键帧、手动加帧，形成参考分镜池。</p></div>
           <div class="step"><div class="num">3</div><h3>清洗水印</h3><p>对关键帧做全图或区域清洗，必要时应用为当前参考图。</p></div>
           <div class="step"><div class="num">4</div><h3>主体识别</h3><p>识别场景和主体候选，只是候选，不应锁死。</p></div>
-          <div class="step"><div class="num">5</div><h3>素材准备</h3><p>清洗关键帧，生成场景图和主体多视角/动作/表情资产包。</p></div>
+          <div class="step"><div class="num">5</div><h3>素材准备</h3><p>清洗关键帧，先生成主体多视角/动作/表情资产包，再生成去主体、相似或换风格场景图。</p></div>
           <div class="step"><div class="num">6</div><h3>分镜改造</h3><p>把参考主体、场景、动作和 SKG 产品放入分镜结构。</p></div>
           <div class="step"><div class="num">7</div><h3>生成视频</h3><p>用分镜 4 图槽、改造目标和时长调用 Seedance / Kling / Veo 3 生视频 API，结果回写到画面工作台节点。</p></div>
           <div class="step"><div class="num">8</div><h3>合成成品</h3><p>片段、字幕、配音、转场合成最终 mp4。当前未实现。</p></div>
@@ -571,7 +571,7 @@
               <tbody>
                 <tr><td><code>web/app/page.tsx</code></td><td>产品工作台主状态：jobs、activeJobId、selectedFrames、clipboard、ReactFlow 节点和边；负责打开/找回画布工作面板。</td></tr>
                 <tr><td><code>web/components/nodes/index.tsx</code></td><td>DAG 节点定义：Input、VisualLab、Audio、Compose，以及画布工作面板 KeyframePanel / VideoFramePanel；旧 Keyframe/Storyboard/VideoGen 组件保留但不再挂主画布。</td></tr>
-                <tr><td><code>web/components/lightbox.tsx</code></td><td>关键帧素材准备面板：清洗、场景图、主体候选、主体资产包和审核。</td></tr>
+                <tr><td><code>web/components/lightbox.tsx</code></td><td>关键帧素材准备面板：清洗、主体候选、主体资产包、去主体场景图和审核。</td></tr>
                 <tr><td><code>web/components/storyboard-bar.tsx</code></td><td>顶部分镜编排条：展示选入编排的关键帧，并作为唯一分镜导航。</td></tr>
                 <tr><td><code>web/components/storyboard-workbench.tsx</code></td><td>顶部分镜编排条下方的明细区：4 图槽、改造目标、时长、自动保存。</td></tr>
                 <tr><td><code>web/lib/api.ts</code></td><td>前端类型和 API client，是前后端数据契约镜像。</td></tr>
@@ -618,13 +618,13 @@ api/main.py
           </div>
           <div class="flow-row">
             <div><strong>你看到的区域</strong><span>画面工作台 · Visual Lab</span></div>
-            <div><strong>主要源码</strong><span><code>VisualLabNode</code> in <code>web/components/nodes/index.tsx</code>；它现在是素材准备看板，汇总关键帧、场景图、主体资产包和视频任务。</span></div>
+            <div><strong>主要源码</strong><span><code>VisualLabNode</code> in <code>web/components/nodes/index.tsx</code>；它现在是素材准备看板，汇总关键帧、主体资产包、场景图和视频任务。</span></div>
             <div><strong>适合怎么描述</strong><span>“画面工作台的素材准备进度、分组缩略图、关键帧审核入口和后续分镜入口应该如何组织”。</span></div>
           </div>
           <div class="flow-row">
             <div><strong>你看到的区域</strong><span>关键帧素材审核面板</span></div>
-            <div><strong>主要源码</strong><span><code>FrameLightbox</code>；按“原图/清洗、场景图、主体资产、审核”四个页签组织；非主体页采用左侧大图 + 右侧窄状态栏，主体资产页保留较宽右栏；清洗页支持一键批量生成待审核清洗版，相关接口包括 <code>cleanupFrame</code>、<code>addElement</code>、<code>generateSceneAsset</code>、<code>generateSubjectAssets</code>。</span></div>
-            <div><strong>适合怎么描述</strong><span>“某张关键帧的水印、场景图、主体多视角/动作/表情图和质量风险应该如何审核”。</span></div>
+            <div><strong>主要源码</strong><span><code>FrameLightbox</code>；按“原图/清洗、主体资产、场景图、审核”四个页签组织；左侧只放主图/框选画布，右侧承载当前页操作、状态和结果；场景图依赖主体资产，支持去主体原场景、相似新场景和同构换风格。相关接口包括 <code>cleanupFrame</code>、<code>addElement</code>、<code>generateSubjectAssets</code>、<code>generateSceneAsset</code>。</span></div>
+            <div><strong>适合怎么描述</strong><span>“某张关键帧的水印、主体多视角/动作/表情图、去主体场景图和质量风险应该如何审核”。</span></div>
           </div>
           <div class="flow-row">
             <div><strong>你看到的区域</strong><span>顶部分镜头编排下拉面板</span></div>
@@ -728,8 +728,8 @@ SubjectAsset {
             <tr><td>应用清洗</td><td><code>POST /cleanup/apply</code></td><td><code>applyCleanedFrame</code></td><td>物理覆盖 frames/{idx}.jpg，并备份原图。</td></tr>
             <tr><td>元素增改删</td><td><code>POST/PATCH/DELETE /elements</code></td><td><code>addElement/updateElement/deleteElement</code></td><td>让用户修正 Vision 错误，避免候选结果锁死。</td></tr>
             <tr><td>元素提取</td><td><code>POST /elements/{element_id}/cutout</code></td><td><code>cutoutElement</code></td><td>调用图像模型生成独立白底素材图，每次累积一张 cutout。</td></tr>
-            <tr><td>场景资产</td><td><code>POST /frames/{idx}/scene-asset</code></td><td><code>generateSceneAsset</code></td><td>为每张已选关键帧生成一张去水印、高清增强的场景图，保留历史版本用于人工审核。</td></tr>
             <tr><td>主体资产包</td><td><code>POST /elements/{element_id}/subject-assets</code></td><td><code>generateSubjectAssets</code></td><td>根据用户选择的视图、动作和表情生成主体资产包；当多个关键帧都指向同一主体时，前端把已选关键帧作为 <code>source_frame_indices</code> 传入，后端拼参考板。</td></tr>
+            <tr><td>场景资产</td><td><code>POST /frames/{idx}/scene-asset</code></td><td><code>generateSceneAsset</code></td><td>在主体资产之后生成去主体背景板；请求包含 <code>scene_mode</code> 和 <code>scene_style</code>，可做原场景补背景、相似新场景或同构换风格，保留历史版本用于人工审核。</td></tr>
             <tr><td>分镜保存</td><td><code>PUT /frames/{idx}/storyboard</code></td><td><code>updateStoryboard</code></td><td>保存 4 图槽、时长和改造说明。</td></tr>
             <tr><td>生图</td><td><code>POST /frames/{idx}/generate</code></td><td><code>generateImage</code></td><td>基于关键帧或已选生成图做 image-to-image，目前可用。</td></tr>
           </tbody>
@@ -751,7 +751,7 @@ SubjectAsset {
             </tr>
             <tr>
               <td><span class="tag violet">画面工作台 Visual Lab</span></td>
-              <td>作为素材准备看板：显示准备进度、质量风险、关键帧 / 场景图 / 主体包 / 分镜视频四个入口；上方缩略图按关键帧、场景图、主体包、视频任务分组。点击关键帧进入素材审核面板，点击资产图复制到分镜编排。</td>
+              <td>作为素材准备看板：显示准备进度、质量风险、关键帧 / 主体包 / 场景图 / 分镜视频四个入口；上方缩略图按关键帧、主体包、场景图、视频任务分组。点击关键帧进入素材审核面板，点击资产图复制到分镜编排。</td>
               <td>不要在主卡片里堆复杂表单；主卡片只做状态总览和入口。</td>
               <td><code>VisualLabNode</code>、<code>FrameLightbox</code>、<code>generateSceneAsset</code>、<code>generateSubjectAssets</code>、视频任务接口</td>
             </tr>
@@ -814,7 +814,7 @@ SubjectAsset {
         <div class="todo">
           <div class="todo-item">
             <h3>改关键帧素材准备</h3>
-            <p>“我在关键帧素材准备面板里，主体候选应该怎么编辑/删除；场景图和主体资产包怎么生成、审核、复制到分镜。”</p>
+            <p>“我在关键帧素材准备面板里，主体候选应该怎么编辑/删除；主体资产包怎么生成；场景图怎么基于主体去除、换风格、审核、复制到分镜。”</p>
           </div>
           <div class="todo-item">
             <h3>改 Storyboard 节点</h3>
@@ -841,13 +841,26 @@ SubjectAsset {
         <div class="changelog">
           <article class="change">
             <header>
-              <h3>2026-05-14 · 关键帧素材面板右侧改为紧凑状态栏</h3>
+              <h3>2026-05-14 · 场景图改为主体资产之后生成</h3>
+              <span class="tag violet">FrameLightbox</span>
+              <span class="tag blue">API</span>
+            </header>
+            <div class="body">
+              <p><strong>问题：</strong>场景图如果先于主体资产生成，只能做普通背景清理，无法准确知道要移除哪个主体，也不利于后续生成相似但不同或同构换风格的场景。</p>
+              <p><strong>改动：</strong><code>FrameLightbox</code> 页签顺序改为“原图/清洗 → 主体资产 → 场景图 → 审核”；画面工作台缩略图和进度文案也同步为主体资产先于场景图。场景图页新增“去主体原场景 / 相似新场景 / 同构换风格”和风格选择，且在没有主体资产时提示先生成主体资产。</p>
+              <p><strong>后端：</strong><code>generateSceneAsset</code> 请求新增 <code>scene_mode</code> 和 <code>scene_style</code>；后端提示词会优先读取已生成主体资产对应的主体名称，生成去主体并补背景的场景图，再按模式决定是否做相似变化或风格变化。</p>
+              <p><strong>影响：</strong><code>web/components/lightbox.tsx</code>、<code>web/components/nodes/index.tsx</code>、<code>web/lib/api.ts</code>、<code>api/main.py</code>、<code>docs/source-analysis.html</code>。</p>
+            </div>
+          </article>
+          <article class="change">
+            <header>
+              <h3>2026-05-14 · 关键帧素材面板统一右侧操作栏</h3>
               <span class="tag violet">FrameLightbox</span>
               <span class="tag blue">Layout</span>
             </header>
             <div class="body">
-              <p><strong>问题：</strong>移除旧元素提取和手工加主体入口后，关键帧详情右侧内容变少，继续占用大列会压缩左侧主图和清洗操作区。</p>
-              <p><strong>改动：</strong><code>FrameLightbox</code> 在“原图/清洗、场景图、审核”页把右侧改成固定窄状态栏，左侧主图和操作区获得更大宽度；“主体资产”页仍保留较宽右栏，用于主体识别、主体清单和资产包。</p>
+              <p><strong>问题：</strong>“原图/清洗、场景图、主体资产、审核”都应遵循同一结构：左侧负责看图和框选，右侧负责操作、状态和结果；旧布局把部分操作塞在左侧下方，导致左侧满、右侧空。</p>
+              <p><strong>改动：</strong><code>FrameLightbox</code> 统一为左侧主图、右侧操作栏。清洗按钮、批量清洗、清洗结果预览、场景图生成/复制、主体识别/主体资产包和审核状态都在右侧；切换到非清洗页时会退出框选模式，避免画框状态残留。</p>
               <p><strong>影响：</strong><code>web/components/lightbox.tsx</code>、<code>docs/source-analysis.html</code>。</p>
             </div>
           </article>

web/components/lightbox.tsx

@@ -6,7 +6,7 @@ import {
   frameUrl, cleanedFrameUrl, apiAssetUrl,
   describeFrame, cleanupFrame, applyCleanedFrame, discardCleanedFrame, addElement, updateElement, deleteElement,
   generateSceneAsset, generateSubjectAssets,
-  type AssetBackground, type AssetSize, type KeyFrame, type Job, type ImageRef, type SubjectKind,
+  type AssetBackground, type AssetSize, type KeyFrame, type Job, type ImageRef, type SceneMode, type SceneStyle, type SubjectKind,
 } from "@/lib/api"
 import { toast } from "sonner"
 
@@ -51,11 +51,25 @@ type LightboxTab = "clean" | "scene" | "subject" | "review"
 
 const LIGHTBOX_TABS: Array<{ key: LightboxTab; label: string }> = [
   { key: "clean", label: "原图/清洗" },
-  { key: "scene", label: "场景图" },
   { key: "subject", label: "主体资产" },
+  { key: "scene", label: "场景图" },
   { key: "review", label: "审核" },
 ]
 
+const SCENE_MODE_OPTIONS: Array<[SceneMode, string]> = [
+  ["remove_subject", "去主体原场景"],
+  ["similar", "相似新场景"],
+  ["style", "同构换风格"],
+]
+
+const SCENE_STYLE_OPTIONS: Array<[SceneStyle, string]> = [
+  ["source", "跟随原图"],
+  ["premium_product", "高端产品感"],
+  ["clean_studio", "干净工作室"],
+  ["warm_lifestyle", "真实生活感"],
+  ["cinematic", "电影感"],
+]
+
 export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, onChange, onToggleSelect, onJobUpdate, onSwitchPanel, onCopyImage, embedded = false }: Props) {
   const [describing, setDescribing] = useState(false)
   const [cleaningFrameIds, setCleaningFrameIds] = useState<Set<number>>(new Set())
@@ -65,6 +79,8 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
   const [sceneGenerating, setSceneGenerating] = useState(false)
   const [subjectGenerating, setSubjectGenerating] = useState<string | null>(null)
   const [assetSize, setAssetSize] = useState<AssetSize>("source")
+  const [sceneMode, setSceneMode] = useState<SceneMode>("remove_subject")
+  const [sceneStyle, setSceneStyle] = useState<SceneStyle>("source")
   const [subjectKinds, setSubjectKinds] = useState<Record<string, SubjectKind>>({})
   const [subjectBackgrounds, setSubjectBackgrounds] = useState<Record<string, AssetBackground>>({})
   const [subjectViews, setSubjectViews] = useState<Record<string, string[]>>({})
@@ -124,6 +140,7 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
   const selectedFrameIndices = Array.from(selected).sort((a, b) => a - b)
   const sharedSubjectFrameIndices = selectedFrameIndices.length > 1 ? selectedFrameIndices : [f.index]
   const subjectAssetCount = elements.reduce((sum, item) => sum + (item.subject_assets?.length ?? 0), 0)
+  const hasSubjectAssets = subjectAssetCount > 0
   const qualityWarnings = [
     ...(f.quality_report?.warnings ?? []),
     ...(latestSceneAsset?.quality_report?.warnings ?? []),
@@ -197,9 +214,14 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
   }
 
   const handleGenerateSceneAsset = async () => {
+    if (!hasSubjectAssets) {
+      setActiveTab("subject")
+      toast.message("先生成主体资产，再生成去主体场景图")
+      return
+    }
     setSceneGenerating(true)
     try {
-      const updated = await generateSceneAsset(jobId, f.index, { size: assetSize })
+      const updated = await generateSceneAsset(jobId, f.index, { size: assetSize, scene_mode: sceneMode, scene_style: sceneStyle })
       onJobUpdate?.(updated)
       toast.success(`分镜 ${f.index + 1} 场景图已生成`)
     } catch (e) {
@@ -406,7 +428,15 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
           <button
             key={tab.key}
             type="button"
-            onClick={() => setActiveTab(tab.key)}
+            onClick={() => {
+              setActiveTab(tab.key)
+              if (tab.key !== "clean") {
+                setCropMode(false)
+                setRegions([])
+                setDraftRegion(null)
+                setDragStart(null)
+              }
+            }}
             className={`h-7 rounded-md px-2.5 text-[11px] font-medium transition ${
               activeTab === tab.key
                 ? "bg-white text-black shadow"
@@ -417,13 +447,13 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
           </button>
         ))}
         <div className="ml-auto hidden items-center gap-2 text-[10px] text-white/42 sm:flex">
-          <span>{latestSceneAsset ? "场景已生成" : "场景待生成"}</span>
+          <span>{hasSubjectAssets ? `${subjectAssetCount} 主体资产` : "主体待生成"}</span>
           <span>·</span>
-          <span>{subjectAssetCount > 0 ? `${subjectAssetCount} 主体资产` : "主体待生成"}</span>
+          <span>{latestSceneAsset ? "场景已生成" : "场景待生成"}</span>
         </div>
       </div>
 
-      {/* 主体 — 左：大图 + 主操作；右：当前页上下文 / 主体资产 */}
+      {/* 主体 — 左：主图；右：当前页操作 / 状态 / 主体资产 */}
       <div className="flex gap-3 p-3 overflow-hidden flex-1 min-h-0">
         {/* 左侧大图区 */}
         <div
@@ -448,11 +478,7 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
               alt={`frame ${f.index}`}
               className="rounded-lg object-contain w-full pointer-events-none"
               style={{
-                maxHeight: isSubjectTab
-                  ? (hasCleaned ? "38vh" : "62vh")
-                  : isCleanTab
-                    ? "68vh"
-                    : (hasCleaned ? "44vh" : "68vh"),
+                maxHeight: isSubjectTab ? "64vh" : "68vh",
               }}
               draggable={false}
             />
@@ -650,11 +676,50 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
                   <option value="2048">2048</option>
                 </select>
               </div>
+              <div className="mb-2 rounded-md border border-white/10 bg-black/25 px-2 py-1.5 text-[10px] leading-relaxed text-white/50">
+                先根据主体资产确认要移除的主体，再补全空场景；之后可生成相似新场景或同构换风格场景。
+              </div>
+              <div className="mb-2 grid grid-cols-2 gap-1.5">
+                <label className="space-y-1">
+                  <span className="block text-[9px] text-white/35">生成方式</span>
+                  <select
+                    value={sceneMode}
+                    onChange={(e) => setSceneMode(e.target.value as SceneMode)}
+                    className="w-full rounded border border-white/10 bg-black/35 px-1.5 py-1 text-[10px] text-white/75 outline-none"
+                  >
+                    {SCENE_MODE_OPTIONS.map(([value, label]) => (
+                      <option key={value} value={value}>{label}</option>
+                    ))}
+                  </select>
+                </label>
+                <label className="space-y-1">
+                  <span className="block text-[9px] text-white/35">风格</span>
+                  <select
+                    value={sceneStyle}
+                    onChange={(e) => setSceneStyle(e.target.value as SceneStyle)}
+                    className="w-full rounded border border-white/10 bg-black/35 px-1.5 py-1 text-[10px] text-white/75 outline-none"
+                  >
+                    {SCENE_STYLE_OPTIONS.map(([value, label]) => (
+                      <option key={value} value={value}>{label}</option>
+                    ))}
+                  </select>
+                </label>
+              </div>
+              {!hasSubjectAssets && (
+                <div className="mb-2 rounded border border-amber-300/25 bg-amber-500/10 px-2 py-1.5 text-[10px] leading-snug text-amber-100/85">
+                  还没有主体资产。先在“主体资产”页生成主体图，场景图才能更准确地去主体和补背景。
+                </div>
+              )}
               {latestSceneAsset ? (
                 <div className="mb-2 overflow-hidden rounded-md border border-emerald-300/25 bg-black/30">
                   <img src={apiAssetUrl(latestSceneAsset.url)} alt={latestSceneAsset.label} className="max-h-44 w-full object-contain bg-black" />
                   <div className="flex items-center justify-between gap-2 border-t border-white/10 px-2 py-1 text-[9.5px] text-white/50">
-                    <span>{latestSceneAsset.width}×{latestSceneAsset.height}</span>
+                    <span>
+                      {latestSceneAsset.width}×{latestSceneAsset.height}
+                      {latestSceneAsset.scene_mode && (
+                        <> · {SCENE_MODE_OPTIONS.find(([value]) => value === latestSceneAsset.scene_mode)?.[1] ?? latestSceneAsset.scene_mode}</>
+                      )}
+                    </span>
                     {onCopyImage && (
                       <button
                         type="button"
@@ -680,13 +745,22 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
               <button
                 type="button"
                 onClick={handleGenerateSceneAsset}
-                disabled={sceneGenerating || isCleaningCurrentFrame || batchCleaning}
+                disabled={sceneGenerating || isCleaningCurrentFrame || batchCleaning || !hasSubjectAssets}
                 className="w-full rounded-md bg-emerald-500/65 px-2 py-1.5 text-[11px] font-medium text-white transition hover:bg-emerald-400 disabled:cursor-wait disabled:opacity-45 inline-flex items-center justify-center gap-1"
-                title="生成一张去水印、高清增强后的场景参考图"
+                title={hasSubjectAssets ? "基于主体资产去主体、补背景并生成场景参考图" : "先生成主体资产"}
               >
                 {sceneGenerating ? <Loader2 className="h-3 w-3 animate-spin" /> : <Sparkles className="h-3 w-3" />}
-                {sceneGenerating ? "生成场景图中…" : latestSceneAsset ? "重新生成场景图" : "生成场景图"}
+                {sceneGenerating ? "生成场景图中…" : latestSceneAsset ? "重新生成场景图" : "生成去主体场景图"}
               </button>
+              {!hasSubjectAssets && (
+                <button
+                  type="button"
+                  onClick={() => setActiveTab("subject")}
+                  className="mt-1.5 w-full rounded-md border border-violet-300/25 bg-violet-500/15 px-2 py-1.5 text-[10.5px] font-medium text-violet-100 transition hover:bg-violet-500/25"
+                >
+                  去生成主体资产
+                </button>
+              )}
             </section>
           )}
           {activeTab === "review" && (
@@ -697,10 +771,6 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
                   <div className="text-[9px] opacity-70">清洗</div>
                   <div>{f.cleaned_applied ? "已应用" : hasCleaned ? "待确认" : "未处理"}</div>
                 </div>
-                <div className={`rounded border px-2 py-1 ${latestSceneAsset ? "border-emerald-300/30 bg-emerald-500/10 text-emerald-100" : "border-white/10 bg-black/25 text-white/55"}`}>
-                  <div className="text-[9px] opacity-70">场景</div>
-                  <div>{latestSceneAsset ? "已生成" : "未生成"}</div>
-                </div>
                 <div className={`rounded border px-2 py-1 ${elements.length > 0 ? "border-violet-300/35 bg-violet-500/12 text-violet-100" : "border-white/10 bg-black/25 text-white/55"}`}>
                   <div className="text-[9px] opacity-70">主体</div>
                   <div>{elements.length} 个</div>
@@ -709,6 +779,10 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
                   <div className="text-[9px] opacity-70">资产</div>
                   <div>{subjectAssetCount} 张</div>
                 </div>
+                <div className={`rounded border px-2 py-1 ${latestSceneAsset ? "border-emerald-300/30 bg-emerald-500/10 text-emerald-100" : "border-white/10 bg-black/25 text-white/55"}`}>
+                  <div className="text-[9px] opacity-70">场景</div>
+                  <div>{latestSceneAsset ? "已生成" : "未生成"}</div>
+                </div>
               </div>
               {qualityWarnings.length > 0 ? (
                 <div className="rounded border border-amber-300/25 bg-amber-500/10 px-2 py-1.5 text-[10px] leading-snug text-amber-100/85">
@@ -722,7 +796,7 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
                 </div>
               )}
               <div className="mt-2 text-[10px] leading-relaxed text-white/42">
-                审核通过后，把场景图和主体资产复制到分镜槽位；当前不会自动覆盖素材。
+                审核通过后，把主体资产和场景图复制到分镜槽位；当前不会自动覆盖素材。
               </div>
             </section>
           )}

web/components/nodes/index.tsx

@@ -1257,18 +1257,6 @@ export function VisualLabNode({ data, selected }: any) {
       borderClass: f.quality_report?.risk === "bad" ? "border-rose-300/70" : f.quality_report?.risk === "warn" ? "border-amber-300/70" : "border-orange-300/50",
       aspect,
     })) : []),
-    ...sceneAssets.map((p) => ({
-      id: `scene:${p.frameIdx}:${p.assetId}`,
-      kind: "scene" as const,
-      group: "场景图",
-      frameIdx: p.frameIdx,
-      assetId: p.assetId,
-      src: p.src,
-      label: p.label,
-      caption: `${p.width}×${p.height}`,
-      borderClass: p.risk === "bad" ? "border-rose-300/70" : p.risk === "warn" ? "border-amber-300/70" : "border-emerald-300/60",
-      aspect: p.width && p.height ? `${p.width}/${p.height}` : aspect,
-    })),
     ...subjectAssets.map((p) => ({
       id: `subject:${p.frameIdx}:${p.assetId}`,
       kind: "subject" as const,
@@ -1281,6 +1269,18 @@ export function VisualLabNode({ data, selected }: any) {
       borderClass: "border-violet-300/65",
       aspect: p.width && p.height ? `${p.width}/${p.height}` : "1/1",
     })),
+    ...sceneAssets.map((p) => ({
+      id: `scene:${p.frameIdx}:${p.assetId}`,
+      kind: "scene" as const,
+      group: "场景图",
+      frameIdx: p.frameIdx,
+      assetId: p.assetId,
+      src: p.src,
+      label: p.label,
+      caption: `${p.width}×${p.height}`,
+      borderClass: p.risk === "bad" ? "border-rose-300/70" : p.risk === "warn" ? "border-amber-300/70" : "border-emerald-300/60",
+      aspect: p.width && p.height ? `${p.width}/${p.height}` : aspect,
+    })),
     ...videos.map((v, i) => {
       const videoSrc = apiAssetUrl(v.url)
       const posterSrc = apiAssetUrl(v.poster_url)
@@ -1556,19 +1556,6 @@ export function VisualLabNode({ data, selected }: any) {
             </div>
             <div>关键帧素材</div>
           </button>
-          <button
-            type="button"
-            onClick={(e) => { e.stopPropagation(); openFirstFrame() }}
-            disabled={!job || frames.length === 0}
-            className="min-h-14 rounded-md border border-white/10 px-2 py-2 text-left transition hover:border-emerald-300/50 hover:bg-emerald-400/10 disabled:opacity-35"
-            title="生成 / 审核每张关键帧的场景图"
-          >
-            <div className="mb-1 flex items-center gap-1 text-[var(--text-strong)] text-[12px] font-semibold">
-              <Sparkles className="h-3 w-3 text-emerald-300" />
-              {sceneAssetCount}/{targetFrameCount}
-            </div>
-            <div>场景图</div>
-          </button>
           <button
             type="button"
             onClick={(e) => { e.stopPropagation(); openFirstFrame() }}
@@ -1582,6 +1569,19 @@ export function VisualLabNode({ data, selected }: any) {
             </div>
             <div>主体资产</div>
           </button>
+          <button
+            type="button"
+            onClick={(e) => { e.stopPropagation(); openFirstFrame() }}
+            disabled={!job || frames.length === 0}
+            className="min-h-14 rounded-md border border-white/10 px-2 py-2 text-left transition hover:border-emerald-300/50 hover:bg-emerald-400/10 disabled:opacity-35"
+            title="基于主体资产生成去主体 / 相似 / 换风格场景图"
+          >
+            <div className="mb-1 flex items-center gap-1 text-[var(--text-strong)] text-[12px] font-semibold">
+              <Sparkles className="h-3 w-3 text-emerald-300" />
+              {sceneAssetCount}/{targetFrameCount}
+            </div>
+            <div>场景图</div>
+          </button>
           <button
             type="button"
             onClick={(e) => { e.stopPropagation(); d.onOpenWorkbench?.(frames.find((f) => d.selectedFrames.has(f.index))?.index ?? frames[0]?.index) }}
@@ -1599,10 +1599,10 @@ export function VisualLabNode({ data, selected }: any) {
         <div className="mt-2 text-[10.5px] leading-snug text-[var(--text-faint)]">
           {frames.length > 0 ? (
             <>
-              {cleanedCount} 已清洗 · {sceneAssetCount} 场景图 · {subjectAssetCount} 主体资产 · {targetFrameCount} 素材帧 · {completedVideos.length} 已完成
+              {cleanedCount} 已清洗 · {subjectAssetCount} 主体资产 · {sceneAssetCount} 场景图 · {targetFrameCount} 素材帧 · {completedVideos.length} 已完成
             </>
           ) : (
-            "解析后这里变成素材准备看板：先审关键帧，再生成场景图和主体资产包。"
+            "解析后这里变成素材准备看板：先审关键帧，再生成主体资产包和去主体场景图。"
           )}
         </div>
       </NodeShell>
@@ -1771,7 +1771,7 @@ export function KeyframeNode({ data, selected }: any) {
               <span className={elementsCount > 0 ? "text-violet-300/90 font-medium" : ""}>{elementsCount} 主体候选</span>
               <br />
               <span className="text-[10.5px] text-[var(--text-faint)]">
-                点缩略图 → 清洗水印 / 准备场景图和主体资产 → 改造成 SKG 画面素材
+                点缩略图 → 清洗水印 / 准备主体资产和去主体场景图 → 改造成 SKG 画面素材
               </span>
             </div>
           )

web/lib/api.ts

@@ -139,6 +139,8 @@ export type AssetBackground = "white" | "black"
 export type AssetSize = "source" | "1024" | "1536" | "2048"
 export type SubjectKind = "object" | "living"
 export type SubjectView = string
+export type SceneMode = "remove_subject" | "similar" | "style"
+export type SceneStyle = "source" | "premium_product" | "clean_studio" | "warm_lifestyle" | "cinematic"
 
 export interface QualityReport {
   width: number
@@ -157,6 +159,8 @@ export interface SceneAsset {
   height: number
   quality: "hd"
   size: AssetSize
+  scene_mode?: SceneMode
+  scene_style?: SceneStyle
   quality_report?: QualityReport | null
   created_at: number
 }
@@ -653,12 +657,17 @@ export async function cutoutElement(jobId: string, frameIdx: number, elementId:
 export async function generateSceneAsset(
   jobId: string,
   frameIdx: number,
-  body: { size?: AssetSize } = {},
+  body: { size?: AssetSize; scene_mode?: SceneMode; scene_style?: SceneStyle } = {},
 ): Promise<Job> {
   const res = await fetch(`${API_BASE}/jobs/${jobId}/frames/${frameIdx}/scene-asset`, {
     method: "POST",
     headers: { "Content-Type": "application/json" },
-    body: JSON.stringify({ quality: "hd", size: body.size ?? "source" }),
+    body: JSON.stringify({
+      quality: "hd",
+      size: body.size ?? "source",
+      scene_mode: body.scene_mode ?? "remove_subject",
+      scene_style: body.scene_style ?? "source",
+    }),
   })
   if (!res.ok) {
     const txt = await res.text().catch(() => "")