auto-save 2026-05-14 06:38 (~3)

.memory/worklog.json

@@ -1,19 +1,5 @@
 {
   "entries": [
-    {
-      "files_changed": 1,
-      "hash": "440164e",
-      "message": "auto-save 2026-05-12 17:50 (~1)",
-      "ts": "2026-05-12T17:51:03+08:00",
-      "type": "commit"
-    },
-    {
-      "files_changed": 2,
-      "hash": "aa5ad08",
-      "message": "auto-save 2026-05-12 18:29 (+1, ~1)",
-      "ts": "2026-05-12T18:29:59+08:00",
-      "type": "commit"
-    },
     {
       "files_changed": 3,
       "hash": "64db093",
@@ -3348,6 +3334,19 @@
       "type": "session-heartbeat",
       "message": "Claude 会话活跃 · 最近命令：claude · 5 项未提交变更 · 最近提交：auto-save 2026-05-14 06:27 (~4)",
       "files_changed": 5
+    },
+    {
+      "ts": "2026-05-14T06:33:37+08:00",
+      "type": "commit",
+      "message": "auto-save 2026-05-14 06:33 (~5)",
+      "hash": "0d86b4c",
+      "files_changed": 5
+    },
+    {
+      "ts": "2026-05-13T22:38:51Z",
+      "type": "session-heartbeat",
+      "message": "Codex 会话活跃 · 最近命令：codex · 3 项未提交变更 · 最近提交：auto-save 2026-05-14 06:33 (~5)",
+      "files_changed": 3
     }
   ]
 }

docs/source-analysis.html

@@ -623,7 +623,7 @@ api/main.py
           </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、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，但主体资产页左侧改为全部参考帧网格；右侧承载当前页操作、状态和结果；主体资产页只确认一个统一主体，默认用全部关键帧或已选关键帧作为参考，后端按参考重绘六张纯背景、占满画面的标准站立主体图；场景图依赖主体资产，按当前关键帧生成去主体原场景、相似新场景或同构换风格。相关接口包括 <code>cleanupFrame</code>、<code>addElement</code>、<code>generateSubjectAssets</code>、<code>generateSceneAsset</code>。</span></div>
+            <div><strong>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，但主体资产页左侧改为全部已清洗/已选参考帧网格，场景图页左侧显示全部关键帧并可勾选场景参考；右侧承载当前页操作、状态和结果。主体资产页只确认一个统一主体，后端按参考重绘六张纯背景、占满画面的标准站立主体图；场景图依赖主体资产，右侧通过地点、生成方式、风格和参考要素拼出可编辑 prompt，再按当前关键帧生成去主体原场景、相似新场景或同构换风格。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>“这一组关键帧如何共同生成一个统一主体包；某张关键帧的水印、去主体场景图和质量风险应该如何审核”。</span></div>
           </div>
           <div class="flow-row">
@@ -731,7 +731,7 @@ SubjectAsset {
             <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>主体资产包</td><td><code>POST /elements/{element_id}/subject-assets</code></td><td><code>generateSubjectAssets</code></td><td>根据参考帧重新绘制一个统一主体资产包；前端默认把全部关键帧作为 <code>source_frame_indices</code>，如果用户手动选择了关键帧则只传已选帧，后端拼参考板。默认输出六张标准站立/转身参考图，纯白/黑背景，不含其他元素，并裁去空白让主体占满画面。</td></tr>
-            <tr><td>场景资产</td><td><code>POST /frames/{idx}/scene-asset</code></td><td><code>generateSceneAsset</code></td><td>在统一主体资产之后，按当前关键帧生成去主体背景板；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>场景资产</td><td><code>POST /frames/{idx}/scene-asset</code></td><td><code>generateSceneAsset</code></td><td>在统一主体资产之后，按当前关键帧生成去主体背景板；请求包含 <code>scene_mode</code>、<code>scene_style</code>、<code>prompt</code> 和 <code>source_frame_indices</code>，可用左侧选择的参考帧 + 右侧关键词生成原场景补背景、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>
@@ -841,6 +841,19 @@ SubjectAsset {
         <h2>变更记录</h2>
         <p>这个记录不是 git log 的替代品。它记录“产品理解发生了什么变化、影响了哪些源码、你以后描述需求时该怎么说”。后续每次改功能都要补一条。</p>
         <div class="changelog">
+          <article class="change">
+            <header>
+              <h3>2026-05-14 · 场景图改为全图参考和关键词 Prompt</h3>
+              <span class="tag violet">FrameLightbox</span>
+              <span class="tag blue">Prompt</span>
+            </header>
+            <div class="body">
+              <p><strong>问题：</strong>场景图页不能只围绕当前单张图；它需要看到全部关键帧，并通过地点、风格、参考要素等关键词组合出可控 prompt，再生成场景。</p>
+              <p><strong>改动：</strong><code>FrameLightbox</code> 的“场景图”页左侧改为全部关键帧网格：点击图片设为生成目标，点击“选”加入场景参考。右侧新增地点、生成方式、风格、额外关键词和参考要素 chips，并自动拼出可编辑场景 prompt。</p>
+              <p><strong>后端：</strong><code>generateSceneAsset</code> 请求新增 <code>prompt</code> 和 <code>source_frame_indices</code>；多张参考帧会拼成 contact sheet 给图像模型，同时把用户 prompt 注入场景生成提示词。</p>
+              <p><strong>影响：</strong><code>web/components/lightbox.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 · 主体资产改为参考重绘六张标准图</h3>
@@ -849,7 +862,7 @@ SubjectAsset {
             </header>
             <div class="body">
               <p><strong>问题：</strong>主体资产不是抠图，也不是只看当前单帧生成多角度；主体页需要看到全部参考帧，并用这些参考重新绘制一个完整主体。</p>
-              <p><strong>改动：</strong><code>FrameLightbox</code> 在“主体资产”页左侧显示全部参考帧网格，小图排列，可点击切换当前帧；右侧仍负责统一主体确认和生成。人物/生物默认视图改为六张标准站立/转身图：正面、背面、左侧、右侧、左前 45°、右前 45°。</p>
+              <p><strong>改动：</strong><code>FrameLightbox</code> 在“主体资产”页左侧显示参考帧网格，优先纳入所有已清洗帧，额外已选帧也会并入；小图排列，可点击切换当前帧。右侧仍负责统一主体确认和生成。人物/生物默认视图改为六张标准站立/转身图：正面、背面、左侧、右侧、左前 45°、右前 45°。</p>
               <p><strong>后端：</strong><code>generateSubjectAssets</code> prompt 改为“参考重绘”，明确禁止裁剪/抠图/粘贴源像素，要求主体完整居中、纯白/黑背景、无其他元素，并占画面约 85-95% 高度；落盘时会裁掉纯背景空白并放大主体。</p>
               <p><strong>影响：</strong><code>web/components/lightbox.tsx</code>、<code>web/components/nodes/index.tsx</code>、<code>api/main.py</code>、<code>docs/source-analysis.html</code>。</p>
             </div>

web/components/lightbox.tsx

@@ -42,6 +42,27 @@ const LIVING_VIEW_OPTIONS = [
   ["three_quarter_right", "右前 45°"],
 ]
 
+const LIVING_EXPRESSION_OPTIONS = [
+  ["expression_neutral", "中性脸"],
+  ["expression_smile", "微笑"],
+  ["expression_happy", "开心"],
+  ["expression_serious", "严肃"],
+  ["expression_surprised", "惊讶"],
+]
+
+const LIVING_ACTION_OPTIONS = [
+  ["action_walk", "走路"],
+  ["action_turn", "转身"],
+  ["action_hold", "手持"],
+  ["action_use", "使用"],
+]
+
+const LIVING_VIEW_GROUPS = [
+  { title: "身份标准图", hint: "默认必出 · 用来锁定长相、体型、服装和比例", options: LIVING_VIEW_OPTIONS },
+  { title: "表情补充", hint: "需要口播、反应或情绪镜头时再勾", options: LIVING_EXPRESSION_OPTIONS },
+  { title: "动作补充", hint: "需要动作镜头时再勾，仍保持同一人物身份", options: LIVING_ACTION_OPTIONS },
+]
+
 type LightboxTab = "clean" | "scene" | "subject" | "review"
 
 const LIGHTBOX_TABS: Array<{ key: LightboxTab; label: string }> = [
@@ -550,6 +571,8 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
           className="flex flex-col items-stretch gap-2 overflow-y-auto pr-1"
           style={isSubjectTab
             ? { flex: "1 1 360px", minWidth: 220, maxWidth: 460, minHeight: 0 }
+            : isSceneTab
+              ? { flex: "1 1 430px", minWidth: 280, maxWidth: 560, minHeight: 0 }
             : isCleanTab
               ? { flex: "1 1 500px", minWidth: 300, maxWidth: 600, minHeight: 0 }
               : { flex: "1 1 560px", minWidth: 300, maxWidth: 680, minHeight: 0 }}
@@ -606,6 +629,76 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
                 这些参考帧会一起传给模型，用来重绘同一个主体；不是逐张抠图。
               </div>
             </section>
+          ) : isSceneTab ? (
+            <section className="rounded-lg border border-emerald-300/15 bg-emerald-500/[0.06] p-2.5">
+              <div className="mb-2 flex items-center justify-between gap-2">
+                <div className="text-[12px] font-semibold text-white">场景参考图</div>
+                <span className="text-[9.5px] font-mono text-white/38">
+                  {selectedFrameIndices.length > 0 ? `${selectedFrameIndices.length} 已选参考` : "默认当前帧"}
+                </span>
+              </div>
+              <div
+                className="grid gap-2"
+                style={{ gridTemplateColumns: "repeat(auto-fill, minmax(104px, 1fr))" }}
+              >
+                {frames.map((frame) => {
+                  const active = frame.index === f.index
+                  const checked = selected.has(frame.index)
+                  return (
+                    <div
+                      key={frame.index}
+                      className={`overflow-hidden rounded-md border bg-black/35 transition ${
+                        active
+                          ? "border-emerald-300/70 shadow-[0_0_0_1px_rgba(110,231,183,0.22)]"
+                          : "border-white/10 hover:border-emerald-300/45"
+                      }`}
+                    >
+                      <button
+                        type="button"
+                        onClick={() => onChange(frame.index)}
+                        className="block w-full text-left"
+                        title={`设为生成目标：分镜 ${frame.index + 1}`}
+                      >
+                        <div className="relative aspect-[9/13] bg-black">
+                          <img
+                            src={referenceFrameSrc(frame)}
+                            alt={`scene reference ${frame.index}`}
+                            className="h-full w-full object-contain"
+                            draggable={false}
+                          />
+                          <span className="absolute left-1 top-1 rounded bg-black/65 px-1 py-0.5 text-[8.5px] font-mono text-white/80">
+                            {String(frame.index + 1).padStart(2, "0")}
+                          </span>
+                          {active && (
+                            <span className="absolute right-1 top-1 rounded bg-emerald-500/80 px-1 py-0.5 text-[8px] text-white">
+                              目标
+                            </span>
+                          )}
+                        </div>
+                      </button>
+                      <div className="flex items-center justify-between gap-1 px-1.5 py-1 text-[9.5px] text-white/52">
+                        <span>{frame.timestamp.toFixed(2)}s</span>
+                        <button
+                          type="button"
+                          onClick={() => onToggleSelect(frame.index)}
+                          className={`rounded px-1.5 py-0.5 transition ${
+                            checked
+                              ? "bg-emerald-400/80 text-black"
+                              : "bg-white/10 text-white/55 hover:bg-white/18 hover:text-white"
+                          }`}
+                          title={checked ? "取消场景参考" : "加入场景参考"}
+                        >
+                          {checked ? "参考" : "选"}
+                        </button>
+                      </div>
+                    </div>
+                  )
+                })}
+              </div>
+              <div className="mt-2 text-[10px] leading-relaxed text-white/38">
+                左侧显示全部关键帧；点图片设为生成目标，点“选”加入场景参考。未选择时默认只参考当前目标帧。
+              </div>
+            </section>
           ) : (
             <div
               ref={imgWrapRef}
@@ -666,7 +759,7 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
         {/* 右侧主体识别 + 主体资产 */}
         <div
           className="flex flex-col gap-2.5 overflow-y-auto min-h-0"
-          style={isSubjectTab
+          style={isSubjectTab || isSceneTab
             ? { flex: "0 0 360px", width: 360, minWidth: 320 }
             : { flex: "0 0 320px", width: 320, minWidth: 280, maxWidth: 340 }}
         >
@@ -817,12 +910,24 @@ 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">
-                先用多张关键帧生成统一主体资产，再按当前关键帧去除主体并补全空场景；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>
+	              <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">
+	                左侧选择场景参考图，右侧选择地点和参考关键词；下方 prompt 可自动拼好，也可以手动改。
+	              </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">地点</span>
+	                  <select
+	                    value={sceneLocation}
+	                    onChange={(e) => setSceneLocation(e.target.value)}
+	                    className="w-full rounded border border-white/10 bg-black/35 px-1.5 py-1 text-[10px] text-white/75 outline-none"
+	                  >
+	                    {SCENE_LOCATION_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">生成方式</span>
                   <select
                     value={sceneMode}
                     onChange={(e) => setSceneMode(e.target.value as SceneMode)}
@@ -830,11 +935,13 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
                   >
                     {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>
+	                </label>
+	              </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">风格</span>
                   <select
                     value={sceneStyle}
                     onChange={(e) => setSceneStyle(e.target.value as SceneStyle)}
@@ -842,11 +949,63 @@ export function FrameLightbox({ jobId, frames, activeIndex, selected, onClose, o
                   >
                     {SCENE_STYLE_OPTIONS.map(([value, label]) => (
                       <option key={value} value={value}>{label}</option>
-                    ))}
-                  </select>
-                </label>
-              </div>
-              {!hasSubjectAssets && (
+	                    ))}
+	                  </select>
+	                </label>
+	                <label className="space-y-1">
+	                  <span className="block text-[9px] text-white/35">额外关键词</span>
+	                  <input
+	                    value={sceneExtraKeywords}
+	                    onChange={(e) => setSceneExtraKeywords(e.target.value)}
+	                    placeholder="例如：玻璃、金属、夜景"
+	                    className="w-full rounded border border-white/10 bg-black/35 px-1.5 py-1 text-[10px] text-white/75 outline-none placeholder:text-white/25"
+	                  />
+	                </label>
+	              </div>
+	              <div className="mb-2 rounded-md border border-white/10 bg-black/20 p-2">
+	                <div className="mb-1 text-[9px] text-white/35">参考要素</div>
+	                <div className="flex flex-wrap gap-1">
+	                  {SCENE_REFERENCE_OPTIONS.map(([value, label]) => {
+	                    const active = sceneReferenceKeys.includes(value)
+	                    return (
+	                      <button
+	                        key={value}
+	                        type="button"
+	                        onClick={() => setSceneReferenceKeys((prev) => (
+	                          prev.includes(value)
+	                            ? prev.filter((item) => item !== value)
+	                            : [...prev, value]
+	                        ))}
+	                        className={`rounded border px-1.5 py-0.5 text-[9.5px] transition ${
+	                          active
+	                            ? "border-emerald-300/60 bg-emerald-500/35 text-white"
+	                            : "border-white/10 bg-black/25 text-white/45 hover:text-white"
+	                        }`}
+	                      >
+	                        {label}
+	                      </button>
+	                    )
+	                  })}
+	                </div>
+	              </div>
+	              <label className="mb-2 block">
+	                <div className="mb-1 flex items-center justify-between gap-2">
+	                  <span className="text-[9px] text-white/35">场景 prompt</span>
+	                  <button
+	                    type="button"
+	                    onClick={() => setScenePrompt(scenePromptDraft)}
+	                    className="rounded bg-white/10 px-1.5 py-0.5 text-[9.5px] text-white/60 hover:bg-white/18 hover:text-white"
+	                  >
+	                    重组
+	                  </button>
+	                </div>
+	                <textarea
+	                  value={scenePrompt || scenePromptDraft}
+	                  onChange={(e) => setScenePrompt(e.target.value)}
+	                  className="h-28 w-full resize-none rounded-md border border-white/10 bg-black/35 px-2 py-1.5 text-[10px] leading-relaxed text-white/75 outline-none"
+	                />
+	              </label>
+	              {!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>