auto-save 2026-05-18 06:33 (~5)

This commit is contained in:
2026-05-18 06:33:24 +08:00
parent 4a5c549f08
commit 18fd8c9508
5 changed files with 247 additions and 64 deletions

View File

@@ -1,51 +1,5 @@
{ {
"entries": [ "entries": [
{
"files_changed": 1,
"message": "Codex 会话活跃 · 最近命令codex · 1 项未提交变更 · 最近提交auto-save 2026-05-15 15:59 (~4)",
"ts": "2026-05-15T08:04:47Z",
"type": "session-heartbeat"
},
{
"files_changed": 1,
"hash": "ac36c4e",
"message": "auto-save 2026-05-15 16:05 (~1)",
"ts": "2026-05-15T16:05:39+08:00",
"type": "commit"
},
{
"files_changed": 5,
"hash": "5ca9846",
"message": "auto-save 2026-05-15 16:10 (+1, ~4)",
"ts": "2026-05-15T16:11:10+08:00",
"type": "commit"
},
{
"files_changed": 1,
"message": "Codex 会话活跃 · 最近命令codex · 1 项未提交变更 · 最近提交auto-save 2026-05-15 16:10 (+1, ~4)",
"ts": "2026-05-15T08:14:47Z",
"type": "session-heartbeat"
},
{
"files_changed": 1,
"hash": "e262285",
"message": "auto-save 2026-05-15 16:16 (~1)",
"ts": "2026-05-15T16:16:43+08:00",
"type": "commit"
},
{
"files_changed": 1,
"hash": "36eb205",
"message": "auto-save 2026-05-15 16:22 (~1)",
"ts": "2026-05-15T16:22:15+08:00",
"type": "commit"
},
{
"files_changed": 1,
"message": "Codex 会话活跃 · 最近命令codex · 1 项未提交变更 · 最近提交auto-save 2026-05-15 16:22 (~1)",
"ts": "2026-05-15T08:24:47Z",
"type": "session-heartbeat"
},
{ {
"files_changed": 3, "files_changed": 3,
"hash": "ecc5894", "hash": "ecc5894",
@@ -3257,6 +3211,50 @@
"message": "auto-save 2026-05-18 01:02 (~2)", "message": "auto-save 2026-05-18 01:02 (~2)",
"hash": "4c43d89", "hash": "4c43d89",
"files_changed": 2 "files_changed": 2
},
{
"ts": "2026-05-18T01:07:51+08:00",
"type": "commit",
"message": "auto-save 2026-05-18 01:07 (~8)",
"hash": "7ca5a95",
"files_changed": 8
},
{
"ts": "2026-05-17T17:08:32Z",
"type": "session-heartbeat",
"message": "Codex 会话活跃 · 最近命令codex · 分支 main · 1 项未提交变更 · 最近提交auto-save 2026-05-18 01:07 (~8)",
"files_changed": 1
},
{
"ts": "2026-05-17T17:18:32Z",
"type": "session-heartbeat",
"message": "Codex 会话活跃 · 最近命令codex · 分支 main · 1 项未提交变更 · 最近提交auto-save 2026-05-18 01:07 (~8)",
"files_changed": 1
},
{
"ts": "2026-05-17T17:28:32Z",
"type": "session-heartbeat",
"message": "Codex 会话活跃 · 最近命令codex · 分支 main · 1 项未提交变更 · 最近提交auto-save 2026-05-18 01:07 (~8)",
"files_changed": 1
},
{
"ts": "2026-05-17T19:28:08Z",
"type": "session-heartbeat",
"message": "Codex 会话活跃 · 最近命令codex · 分支 main · 1 项未提交变更 · 最近提交auto-save 2026-05-18 01:07 (~8)",
"files_changed": 1
},
{
"ts": "2026-05-18T06:22:31+08:00",
"type": "commit",
"message": "fix: force gpt image model",
"hash": "4a5c549",
"files_changed": 0
},
{
"ts": "2026-05-17T22:23:44Z",
"type": "session-heartbeat",
"message": "Codex 会话活跃 · 最近命令codex · 分支 main · 1 项未提交变更 · 最近提交fix: force gpt image model",
"files_changed": 1
} }
] ]
} }

View File

@@ -1134,6 +1134,17 @@ def _asset_url(job_id: str, asset_id: str) -> str:
return f"/jobs/{job_id}/assets/{asset_id}.jpg" return f"/jobs/{job_id}/assets/{asset_id}.jpg"
def _delete_subject_asset_file(job_id: str, asset_id: str) -> None:
if not asset_id:
return
p = job_dir(job_id) / "assets" / f"{asset_id}.jpg"
if p.exists():
try:
p.unlink()
except OSError:
pass
def _find_frame(job: Job, idx: int) -> KeyFrame: def _find_frame(job: Job, idx: int) -> KeyFrame:
frame = next((f for f in job.frames if f.index == idx), None) frame = next((f for f in job.frames if f.index == idx), None)
if not frame: if not frame:
@@ -3560,6 +3571,7 @@ class GenerateSubjectAssetsReq(BaseModel):
subject_style: Literal["transparent_human", "source_actor"] = "transparent_human" subject_style: Literal["transparent_human", "source_actor"] = "transparent_human"
reconstruction_mode: Literal["same", "similar"] = "same" reconstruction_mode: Literal["same", "similar"] = "same"
prompt: str = "" prompt: str = ""
replace_views: bool = False
class UpdateProductRefsReq(BaseModel): class UpdateProductRefsReq(BaseModel):
@@ -3974,6 +3986,17 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
if sheet: if sheet:
model_src = sheet model_src = sheet
try:
with Image.open(_source_frame_path(job_id, idx)) as src_im:
source_is_portrait = src_im.height > src_im.width
except Exception:
source_is_portrait = False
canvas_clause = (
"Canvas and aspect ratio: the reference video frame is vertical, so output a vertical portrait 9:16-style image, not a square canvas and not a horizontal layout. "
if source_is_portrait
else "Canvas and aspect ratio: keep a single clean reference-image canvas with the same broad orientation as the source evidence. "
)
target = (el.name_en or el.name_zh).strip() target = (el.name_en or el.name_zh).strip()
bg_phrase = "pure white" if req.background == "white" else "pure black" bg_phrase = "pure white" if req.background == "white" else "pure black"
similar_actor = req.subject_kind == "living" and req.subject_style == "source_actor" and req.reconstruction_mode == "similar" similar_actor = req.subject_kind == "living" and req.subject_style == "source_actor" and req.reconstruction_mode == "similar"
@@ -4036,6 +4059,7 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
+ single_view_clause + single_view_clause
+ identity_clause + identity_clause
+ identity_lock_clause + identity_lock_clause
+ canvas_clause
+ prompt_extra_clause + prompt_extra_clause
+ actor_style_clause + actor_style_clause
+ "The subject must be complete, centered, full body or full object, head-to-feet visible when applicable, not cropped by the canvas. " + "The subject must be complete, centered, full body or full object, head-to-feet visible when applicable, not cropped by the canvas. "
@@ -4082,12 +4106,45 @@ def generate_subject_assets(job_id: str, idx: int, element_id: str, req: Generat
if e.id == element_id: if e.id == element_id:
e.subject_kind = req.subject_kind e.subject_kind = req.subject_kind
e.cutout_background = req.background e.cutout_background = req.background
e.subject_assets = (e.subject_assets or []) + generated current_assets = e.subject_assets or []
if req.replace_views:
replaced_views = {asset.view for asset in generated}
for old_asset in current_assets:
if old_asset.view in replaced_views:
_delete_subject_asset_file(job_id, old_asset.id)
current_assets = [asset for asset in current_assets if asset.view not in replaced_views]
e.subject_assets = current_assets + generated
new_frames.append(f) new_frames.append(f)
update(job, frames=new_frames, message=f"主体资产包生成完成 · {el.name_zh} · {len(generated)}") update(job, frames=new_frames, message=f"主体资产包生成完成 · {el.name_zh} · {len(generated)}")
return job return job
@app.delete("/jobs/{job_id}/frames/{idx}/elements/{element_id}/subject-assets/{asset_id}", response_model=Job)
def delete_subject_asset(job_id: str, idx: int, element_id: str, asset_id: str) -> Job:
"""删除某张主体白底视图。"""
job = JOBS.get(job_id)
if not job:
raise HTTPException(404, "job not found")
frame = _find_frame(job, idx)
el = next((e for e in frame.elements if e.id == element_id), None)
if not el:
raise HTTPException(404, "element not found")
assets = el.subject_assets or []
if not any(asset.id == asset_id for asset in assets):
raise HTTPException(404, "subject asset not found")
_delete_subject_asset_file(job_id, asset_id)
new_frames = []
for f in job.frames:
if f.index == idx:
for e in f.elements:
if e.id == element_id:
e.subject_assets = [asset for asset in (e.subject_assets or []) if asset.id != asset_id]
new_frames.append(f)
update(job, frames=new_frames, message=f"主体视图已删除 · {el.name_zh}")
return job
@app.delete("/jobs/{job_id}/frames/{idx}/elements/{element_id}/cutouts/{cutout_id}", response_model=Job) @app.delete("/jobs/{job_id}/frames/{idx}/elements/{element_id}/cutouts/{cutout_id}", response_model=Job)
def delete_cutout(job_id: str, idx: int, element_id: str, cutout_id: str) -> Job: def delete_cutout(job_id: str, idx: int, element_id: str, cutout_id: str) -> Job:
"""删除该元素的某张提取图""" """删除该元素的某张提取图"""

View File

@@ -589,7 +589,7 @@
<tr><td><code>web/next.config.mjs</code></td><td>Next.js 构建配置:静态导出、图片不走优化、禁用开发环境左下角 Next Dev Indicator并移除 Next 16 已不支持的 <code>eslint</code> 顶层配置,避免本地 dev 出现配置 Issue 提示。</td></tr> <tr><td><code>web/next.config.mjs</code></td><td>Next.js 构建配置:静态导出、图片不走优化、禁用开发环境左下角 Next Dev Indicator并移除 Next 16 已不支持的 <code>eslint</code> 顶层配置,避免本地 dev 出现配置 Issue 提示。</td></tr>
<tr><td><code>web/app/globals.css</code></td><td>全局主题变量、登录页视觉样式、ReactFlow 样式引用,以及本地开发态 <code>nextjs-portal</code> 遮挡隐藏规则。</td></tr> <tr><td><code>web/app/globals.css</code></td><td>全局主题变量、登录页视觉样式、ReactFlow 样式引用,以及本地开发态 <code>nextjs-portal</code> 遮挡隐藏规则。</td></tr>
<tr><td><code>web/app/page.tsx</code></td><td>产品工作台主状态jobs、activeJobId、生成任务状态主渲染为全屏素材输入列 + 信息流广告复刻工作表;“开始”编排状态只负责在下载完成后自动触发 <code>triggerTranscribe</code>不再默认触发抽帧、Vision 扫描或分镜初稿保存;底部吸附音频条不再从主界面渲染。</td></tr> <tr><td><code>web/app/page.tsx</code></td><td>产品工作台主状态jobs、activeJobId、生成任务状态主渲染为全屏素材输入列 + 信息流广告复刻工作表;“开始”编排状态只负责在下载完成后自动触发 <code>triggerTranscribe</code>不再默认触发抽帧、Vision 扫描或分镜初稿保存;底部吸附音频条不再从主界面渲染。</td></tr>
<tr><td><code>web/components/ad-recreation-board.tsx</code></td><td>信息流广告复刻工作表:左侧素材输入只负责链接/上传和任务切换,不再重复放横版原视频预览;右侧展示视频下载状态、默认折叠的文案依据,以及源视频工作区。音频解析结果改成默认折叠的辅助信息,展开后同一行看讲话人/节奏/背景音;主工作区左侧是按 9:16 显示的竖版原视频播放器,播放器内覆盖“当前点抽帧”,按当前播放秒数手动补参考帧;右侧上方是音频波形 / 切点参考,下方是逐句时间轴;下一行铺开“关键帧 / 相似主体”。音频波形用参考图式的连续灰色包络显示响度、停顿和密集爆点,顶部同时显示当前播放秒数、总时长和鼠标指针停点秒数。视频播放时通过 <code>requestAnimationFrame</code> 平滑驱动波形播放线,同时同步高亮并滚动当前句;点击音频波形或字幕行会跳转原视频时间。关键帧区的主入口是“自动抽帧 12 张”,一键按动作峰值目标重新抽取 12 张源视频参考帧,优先抓手势、表情变化、节奏点和镜头变化,缩略图按竖版完整比例显示不裁切并用更多列紧凑铺开,鼠标停留会通过固定浮层放大展示完整帧。“生成 6 视图”放在相似主体白底视图区,不和抽参考按钮平齐;如果用户没有勾选帧,默认把全部关键帧作为主体参考,勾选后只传已选帧;生成区可在“透明骨架 / 普通真人”之间切换,并可填写统一主体方向,例如年轻女性、更运动、更高级。前端调用 <code>generateSubjectAssets</code> 时按主体类型传 <code>subject_style=transparent_human</code><code>source_actor</code>,均使用 <code>reconstruction_mode=similar</code>;后端会把这些帧视为同一个主体的证据,并锁定同一性别表现、年龄段、体型、材质、风格和视觉身份,避免六视图出现男女性别、老少年龄或样式混杂。音频结果下方是信息流复刻分镜工作台:顶部产品参考区是“同一产品素材池”,不限量上传产品图,不做不同产品身份判断;上传原图推荐长边 1200-2000px、短边至少 600px但后端会统一生成最长边 1600px、JPEG 92 的 AI 工作副本,并回显尺寸、自动转换和风险标注;上传后按“套在脖子上的 U 形肩颈按摩仪”进行同一产品批量识别,左/右按佩戴者身体左右、上/下按佩戴方向,额外标注内外侧、开口方向、局部结构点、背景类型、用途标签、生成风险和备注,用户只检查备注,鼠标悬停通过固定浮层显示大图预览,能盖过滚动容器和分镜框架;缺视角补图失败时保留重试入口。脚本区在分镜行上方提供“作者想法”和“整片改写”,每行新口播文案可直接编辑并可单段 AI 改写,分镜时间和原内容列压缩为窄摘要列,把横向空间留给新口播、画面规划和视频候选;生成本条视频时使用当前编辑后的新口播文案。每条音频分镜纵向排列,行内从左到右串起原内容、新口播文案、画面规划/产品融入和 6 个候选视频槽;候选视频槽在宽屏下一排显示 6 个竖版预览,避免前面空旷、后面拥挤。单条生成会从全局选中关键帧或 12 张关键帧中取最贴近本句时间点的参考帧。单条生成会从产品素材池按分镜角色、视角优先级、用途标签、置信度和风险自动挑选最多 6 张相关产品图,不会把全部产品图提交给生视频模型,然后把产品坐标系、视角标注、方向、结构点和风险写入 Seedance 提示。<code>ModelTrace</code> 会在音频解析、产品识别/补图、相似主体 6 视图、脚本改写和单条生视频入口旁直接展示模型名;所有生图入口都显示并使用 <code>gpt-image-2</code>,没有其他图片模型 fallback点击后用固定浮层展示模型链路、输入输出和回退逻辑。旧分镜卡、抽帧控制和视频生成组件仍保留在文件里但当前主路径不渲染。</td></tr> <tr><td><code>web/components/ad-recreation-board.tsx</code></td><td>信息流广告复刻工作表:左侧素材输入只负责链接/上传和任务切换,不再重复放横版原视频预览;右侧展示视频下载状态、默认折叠的文案依据,以及源视频工作区。音频解析结果改成默认折叠的辅助信息,展开后同一行看讲话人/节奏/背景音;主工作区左侧是按 9:16 显示的竖版原视频播放器,播放器内覆盖“当前点抽帧”,按当前播放秒数手动补参考帧;右侧上方是音频波形 / 切点参考,下方是逐句时间轴;下一行铺开“关键帧 / 相似主体”。音频波形用参考图式的连续灰色包络显示响度、停顿和密集爆点,顶部同时显示当前播放秒数、总时长和鼠标指针停点秒数。视频播放时通过 <code>requestAnimationFrame</code> 平滑驱动波形播放线,同时同步高亮并滚动当前句;点击音频波形或字幕行会跳转原视频时间。关键帧区的主入口是“自动抽帧 12 张”,一键按动作峰值目标重新抽取 12 张源视频参考帧,优先抓手势、表情变化、节奏点和镜头变化,缩略图按竖版完整比例显示不裁切并用更多列紧凑铺开,鼠标停留会通过固定浮层放大展示完整帧。“生成 6 视图”放在相似主体白底视图区,不和抽参考按钮平齐;如果用户没有勾选帧,默认把全部关键帧作为主体参考,勾选后只传已选帧;生成区可在“透明骨架 / 普通真人”之间切换,并可填写统一主体方向,例如年轻女性、更运动、更高级。关键帧和相似主体白底视图都用更小的竖版缩略图密排;白底视图只展示每个 view 的最新一张,缩略图上提供“重新生成这一张”和“删除这一张”,单张重生会用 <code>replace_views=true</code> 替换同一视角,不追加成第 7 张。前端调用 <code>generateSubjectAssets</code> 时按主体类型传 <code>subject_style=transparent_human</code><code>source_actor</code>,均使用 <code>reconstruction_mode=similar</code>;后端会把这些帧视为同一个主体的证据,并锁定同一性别表现、年龄段、体型、材质、风格和视觉身份,避免六视图出现男女性别、老少年龄或样式混杂。音频结果下方是信息流复刻分镜工作台:顶部产品参考区是“同一产品素材池”,不限量上传产品图,不做不同产品身份判断;上传原图推荐长边 1200-2000px、短边至少 600px但后端会统一生成最长边 1600px、JPEG 92 的 AI 工作副本,并回显尺寸、自动转换和风险标注;上传后按“套在脖子上的 U 形肩颈按摩仪”进行同一产品批量识别,左/右按佩戴者身体左右、上/下按佩戴方向,额外标注内外侧、开口方向、局部结构点、背景类型、用途标签、生成风险和备注,用户只检查备注,鼠标悬停通过固定浮层显示大图预览,能盖过滚动容器和分镜框架;缺视角补图失败时保留重试入口。脚本区在分镜行上方提供“作者想法”和“整片改写”,每行新口播文案可直接编辑并可单段 AI 改写,分镜时间和原内容列压缩为窄摘要列,把横向空间留给新口播、画面规划和视频候选;生成本条视频时使用当前编辑后的新口播文案。每条音频分镜纵向排列,行内从左到右串起原内容、新口播文案、画面规划/产品融入和 6 个候选视频槽;候选视频槽在宽屏下一排显示 6 个竖版预览,避免前面空旷、后面拥挤。单条生成会从全局选中关键帧或 12 张关键帧中取最贴近本句时间点的参考帧。单条生成会从产品素材池按分镜角色、视角优先级、用途标签、置信度和风险自动挑选最多 6 张相关产品图,不会把全部产品图提交给生视频模型,然后把产品坐标系、视角标注、方向、结构点和风险写入 Seedance 提示。<code>ModelTrace</code> 会在音频解析、产品识别/补图、相似主体 6 视图、脚本改写和单条生视频入口旁直接展示模型名;所有生图入口都显示并使用 <code>gpt-image-2</code>,没有其他图片模型 fallback点击后用固定浮层展示模型链路、输入输出和回退逻辑。旧分镜卡、抽帧控制和视频生成组件仍保留在文件里但当前主路径不渲染。</td></tr>
<tr><td><code>web/app/login/page.tsx</code></td><td>生产登录页:访问账号/访问密钥表单、保持登录、错误/成功状态;当前只在原版 Digital Oasis 动态背景上叠加一个组合登录框,桌面端左侧是动态角色,右侧是图标化登录表单;面板左上角展示官网 SKG 字标和中文“营销内容工作台”系统标识。</td></tr> <tr><td><code>web/app/login/page.tsx</code></td><td>生产登录页:访问账号/访问密钥表单、保持登录、错误/成功状态;当前只在原版 Digital Oasis 动态背景上叠加一个组合登录框,桌面端左侧是动态角色,右侧是图标化登录表单;面板左上角展示官网 SKG 字标和中文“营销内容工作台”系统标识。</td></tr>
<tr><td><code>web/app/login/layout.tsx</code></td><td>登录路由专属 layout覆盖全站默认网页标题和描述为空避免 <code>/login</code> 继承工作台 metadata 后在页面源码里继续出现登录界面文字以外的文案。</td></tr> <tr><td><code>web/app/login/layout.tsx</code></td><td>登录路由专属 layout覆盖全站默认网页标题和描述为空避免 <code>/login</code> 继承工作台 metadata 后在页面源码里继续出现登录界面文字以外的文案。</td></tr>
<tr><td><code>web/components/login/oasis-canvas.tsx</code></td><td>登录页全屏动态视觉层:用 iframe 直接承载下载包 <code>web/public/oasis-source/index.html</code> 的原 WebGPU / Three.js 草场源码;父级登录页只覆盖自己的文案和表单,并在捕获阶段把全局鼠标坐标同时用原生事件和 <code>postMessage</code> 转发给 iframe避免登录面板或输入框遮挡时草地失去鼠标响应。</td></tr> <tr><td><code>web/components/login/oasis-canvas.tsx</code></td><td>登录页全屏动态视觉层:用 iframe 直接承载下载包 <code>web/public/oasis-source/index.html</code> 的原 WebGPU / Three.js 草场源码;父级登录页只覆盖自己的文案和表单,并在捕获阶段把全局鼠标坐标同时用原生事件和 <code>postMessage</code> 转发给 iframe避免登录面板或输入框遮挡时草地失去鼠标响应。</td></tr>
@@ -888,7 +888,7 @@ ProductRefStateItem {
<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>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>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。当前源视频工作区支持 <code>subject_style=transparent_human</code><code>subject_style=source_actor</code> 两种相似主体:透明骨架人会保持透明/半透明皮肤包裹可见白色骨架,普通真人会保持正常广告演员方向。两种模式都使用 <code>reconstruction_mode=similar</code>,最多读取 12 张参考帧,生成 6 张白底视图;后端强制使用 <code>gpt-image-2</code>,不再接受前端或环境变量切到其他图片模型,也不做 <code>gpt-image-1.5</code> / Gemini fallback后端会加身份锁定约束统一性别表现、年龄段、体型、材质、风格和视觉身份避免六视图混成不同人物。前端白底视图缩略图和关键帧一样鼠标停留会用顶层浮层放大预览点击仍打开原图后端每个 <code>view</code> 单独调用一次生图并明确禁止六视图拼图、contact sheet、多主体、多面板、标签或对比排版保证一个视角一张照片。</td></tr> <tr><td>主体资产包</td><td><code>POST /elements/{element_id}/subject-assets</code><br><code>DELETE /elements/{element_id}/subject-assets/{asset_id}</code></td><td><code>generateSubjectAssets</code><br><code>deleteSubjectAsset</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。当前源视频工作区支持 <code>subject_style=transparent_human</code><code>subject_style=source_actor</code> 两种相似主体:透明骨架人会保持透明/半透明皮肤包裹可见白色骨架,普通真人会保持正常广告演员方向。两种模式都使用 <code>reconstruction_mode=similar</code>,最多读取 12 张参考帧,生成 6 张白底视图;后端强制使用 <code>gpt-image-2</code>,不再接受前端或环境变量切到其他图片模型,也不做 <code>gpt-image-1.5</code> / Gemini fallback后端会加身份锁定约束统一性别表现、年龄段、体型、材质、风格和视觉身份避免六视图混成不同人物。如果参考帧是竖屏prompt 会明确要求竖版 9:16 风格画布,落盘也按源帧纵横比归一化。前端白底视图缩略图和关键帧一样,鼠标停留会用顶层浮层放大预览,点击仍打开原图;后端每个 <code>view</code> 单独调用一次生图并明确禁止六视图拼图、contact sheet、多主体、多面板、标签或对比排版保证一个视角一张照片。<code>replace_views=true</code> 时会替换同一视角旧图;删除接口会移除对应 subject asset 记录并删除本地 jpg 文件。</td></tr>
<tr><td>首尾帧资产</td><td><code>POST /frames/{idx}/scene-asset</code></td><td><code>generateSceneAsset</code></td><td>同一接口兼容旧场景图和新首尾帧;新流程传 <code>asset_role=first_frame/last_frame</code>,后端走文字生图,参考帧只用于理解透明骨架人形象、比例、机位和光线,生成结果仍保存在 <code>scene_assets</code> 并自动填入产品融合镜头。</td></tr> <tr><td>首尾帧资产</td><td><code>POST /frames/{idx}/scene-asset</code></td><td><code>generateSceneAsset</code></td><td>同一接口兼容旧场景图和新首尾帧;新流程传 <code>asset_role=first_frame/last_frame</code>,后端走文字生图,参考帧只用于理解透明骨架人形象、比例、机位和光线,生成结果仍保存在 <code>scene_assets</code> 并自动填入产品融合镜头。</td></tr>
<tr><td>产品图库</td><td><code>GET /product-library/skg</code></td><td><code>listProductLibrary</code></td><td>读取内置 SKG 白底图库 manifest返回产品标题、品类、尺寸、白底评分和预览图 URL。</td></tr> <tr><td>产品图库</td><td><code>GET /product-library/skg</code></td><td><code>listProductLibrary</code></td><td>读取内置 SKG 白底图库 manifest返回产品标题、品类、尺寸、白底评分和预览图 URL。</td></tr>
<tr><td>产品图入库到 job</td><td><code>POST /jobs/{id}/assets</code><code>POST /jobs/{id}/assets/product-library</code></td><td><code>uploadStoryboardAsset</code><code>copyProductLibraryAsset</code></td><td>上传产品图或把内置产品图库条目复制为当前 job 的普通 asset。后端统一生成最长边 1600px、JPEG 92 的 AI 工作副本,透明底铺白,过大/过小图片会在 <code>ImageRef.asset_meta</code> 里返回转换动作和风险;黑底/白底背景本身不强行转换。注意该接口只写图片文件,产品素材池列表另由 <code>PUT /jobs/{id}/product-refs</code> 持久化。</td></tr> <tr><td>产品图入库到 job</td><td><code>POST /jobs/{id}/assets</code><code>POST /jobs/{id}/assets/product-library</code></td><td><code>uploadStoryboardAsset</code><code>copyProductLibraryAsset</code></td><td>上传产品图或把内置产品图库条目复制为当前 job 的普通 asset。后端统一生成最长边 1600px、JPEG 92 的 AI 工作副本,透明底铺白,过大/过小图片会在 <code>ImageRef.asset_meta</code> 里返回转换动作和风险;黑底/白底背景本身不强行转换。注意该接口只写图片文件,产品素材池列表另由 <code>PUT /jobs/{id}/product-refs</code> 持久化。</td></tr>
@@ -1004,6 +1004,19 @@ ProductRefStateItem {
<h2>变更记录</h2> <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> <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"> <div class="changelog">
<article class="change">
<header>
<h3>2026-05-18 · 相似主体缩略图压缩并支持单张重生/删除</h3>
<span class="tag rose">UI</span>
<span class="tag violet">API</span>
<span class="tag cyan">Workflow</span>
</header>
<div class="body">
<p><strong>问题:</strong>关键帧和相似主体白底视图缩略图仍偏大;主体 6 视图只能整组重生,某一张不准时无法单独替换;白底视图也缺少删除入口。竖版视频参考生成的主体图还需要明确保持竖版画布。</p>
<p><strong>改动:</strong><code>SourceReferenceBuildPanel</code> 把关键帧网格改成更小的自适应列,把白底主体视图改成 9:16 固定小缩略图;每个主体视图新增重生和删除按钮。前端只展示每个 view 的最新一张;单张重生调用 <code>generateSubjectAssets</code> 并传 <code>views=[当前 view]</code><code>replace_views=true</code>,不会追加成第 7 张。后端 <code>GenerateSubjectAssetsReq</code> 新增 <code>replace_views</code>,并新增 <code>DELETE /subject-assets/{asset_id}</code>;竖屏参考帧会把 9:16 竖版画布要求写入 prompt输出继续按源帧纵横比归一化。</p>
<p><strong>影响:</strong><code>web/components/ad-recreation-board.tsx</code><code>web/lib/api.ts</code><code>api/main.py</code><code>docs/source-analysis.html</code>。后续“重新生成某张主体图”应走同视角替换,不要追加多套六视图。</p>
</div>
</article>
<article class="change"> <article class="change">
<header> <header>
<h3>2026-05-18 · 生图模型统一锁定为 gpt-image-2</h3> <h3>2026-05-18 · 生图模型统一锁定为 gpt-image-2</h3>

View File

@@ -4,7 +4,7 @@ import { type MouseEvent as ReactMouseEvent, type ReactNode, type RefObject, use
import { createPortal } from "react-dom" import { createPortal } from "react-dom"
import { import {
AlertTriangle, Check, ChevronDown, Circle, Film, FileText, Image as ImageIcon, Info, Link2, Loader2, AlertTriangle, Check, ChevronDown, Circle, Film, FileText, Image as ImageIcon, Info, Link2, Loader2,
Mic, Package, PanelRight, Play, Plus, Scissors, Sparkles, Trash2, Upload, Wand2, Mic, Package, PanelRight, Play, Plus, RefreshCw, Scissors, Sparkles, Trash2, Upload, Wand2,
} from "lucide-react" } from "lucide-react"
import { toast } from "sonner" import { toast } from "sonner"
import { import {
@@ -28,6 +28,7 @@ import {
analyzeProductViews, analyzeProductViews,
apiAssetUrl, apiAssetUrl,
cutoutElement, cutoutElement,
deleteSubjectAsset,
effectiveFrameUrl, effectiveFrameUrl,
generateProductAngleAsset, generateProductAngleAsset,
generateSubjectAssets, generateSubjectAssets,
@@ -97,6 +98,8 @@ type AudioStoryboardRow = {
type ProductRefItem = ProductRefStateItem type ProductRefItem = ProductRefStateItem
type SubjectStyleMode = "transparent_human" | "source_actor" type SubjectStyleMode = "transparent_human" | "source_actor"
const SUBJECT_VIEW_ORDER = ["front", "three_quarter_left", "left", "back", "right", "three_quarter_right"]
type ModelTraceSpec = { type ModelTraceSpec = {
title: string title: string
model: string model: string
@@ -1435,6 +1438,7 @@ function SourceReferenceBuildPanel({
}) { }) {
const [extracting, setExtracting] = useState(false) const [extracting, setExtracting] = useState(false)
const [subjectBusy, setSubjectBusy] = useState(false) const [subjectBusy, setSubjectBusy] = useState(false)
const [subjectAssetBusy, setSubjectAssetBusy] = useState<string | null>(null)
const [deletingFrame, setDeletingFrame] = useState<number | null>(null) const [deletingFrame, setDeletingFrame] = useState<number | null>(null)
const [framePreview, setFramePreview] = useState<{ index: number; left: number; top: number } | null>(null) const [framePreview, setFramePreview] = useState<{ index: number; left: number; top: number } | null>(null)
const [subjectAssetPreview, setSubjectAssetPreview] = useState<{ id: string; left: number; top: number } | null>(null) const [subjectAssetPreview, setSubjectAssetPreview] = useState<{ id: string; left: number; top: number } | null>(null)
@@ -1453,8 +1457,22 @@ function SourceReferenceBuildPanel({
return findSimilarActorSource(subjectReferenceFrames, frames) return findSimilarActorSource(subjectReferenceFrames, frames)
}, [frames, subjectReferenceFrames]) }, [frames, subjectReferenceFrames])
const actorAssets = actorSource?.element.subject_assets ?? [] const actorAssets = actorSource?.element.subject_assets ?? []
const visibleActorAssets = useMemo(() => {
const latestByView = new Map<string, SubjectAsset>()
for (const asset of actorAssets) {
const current = latestByView.get(asset.view)
if (!current || (asset.created_at || 0) >= (current.created_at || 0)) {
latestByView.set(asset.view, asset)
}
}
return [...latestByView.values()].sort((a, b) => {
const ai = SUBJECT_VIEW_ORDER.indexOf(a.view)
const bi = SUBJECT_VIEW_ORDER.indexOf(b.view)
return (ai === -1 ? 99 : ai) - (bi === -1 ? 99 : bi)
})
}, [actorAssets])
const previewFrame = framePreview ? frames.find((frame) => frame.index === framePreview.index) ?? null : null const previewFrame = framePreview ? frames.find((frame) => frame.index === framePreview.index) ?? null : null
const previewSubjectAsset = subjectAssetPreview ? actorAssets.find((asset) => asset.id === subjectAssetPreview.id) ?? null : null const previewSubjectAsset = subjectAssetPreview ? visibleActorAssets.find((asset) => asset.id === subjectAssetPreview.id) ?? null : null
const referenceCountLabel = selectedReferenceFrames.length const referenceCountLabel = selectedReferenceFrames.length
? `使用已选 ${selectedReferenceFrames.length}` ? `使用已选 ${selectedReferenceFrames.length}`
: frames.length : frames.length
@@ -1512,6 +1530,7 @@ function SourceReferenceBuildPanel({
source_frame_indices: subjectReferenceFrames.slice(0, 12).map((frame) => frame.index), source_frame_indices: subjectReferenceFrames.slice(0, 12).map((frame) => frame.index),
views: ["front", "back", "left", "right", "three_quarter_left", "three_quarter_right"], views: ["front", "back", "left", "right", "three_quarter_left", "three_quarter_right"],
prompt: buildSimilarSubjectPrompt(subjectStyle, subjectDirection), prompt: buildSimilarSubjectPrompt(subjectStyle, subjectDirection),
replace_views: true,
}) })
onJobUpdate(updated) onJobUpdate(updated)
toast.success("相似主体 6 张白底视图已生成") toast.success("相似主体 6 张白底视图已生成")
@@ -1532,6 +1551,47 @@ function SourceReferenceBuildPanel({
} }
} }
const regenerateSubjectAsset = async (asset: SubjectAsset) => {
if (!actorSource) return
setSubjectAssetBusy(`regen:${asset.id}`)
try {
const sourceIndices = asset.source_frame_indices?.length
? asset.source_frame_indices
: subjectReferenceFrames.slice(0, 12).map((frame) => frame.index)
const updated = await generateSubjectAssets(job.id, actorSource.frame.index, actorSource.element.id, {
subject_kind: "living",
subject_style: subjectStyle,
reconstruction_mode: "similar",
background: asset.background || "white",
size: asset.size || "1024",
source_frame_indices: sourceIndices,
views: [asset.view],
prompt: buildSimilarSubjectPrompt(subjectStyle, subjectDirection),
replace_views: true,
})
onJobUpdate(updated)
toast.success("已重新生成这张主体视图")
} catch (e) {
toast.error("单张主体视图重生失败:" + (e instanceof Error ? e.message : String(e)))
} finally {
setSubjectAssetBusy(null)
}
}
const deleteActorAsset = async (asset: SubjectAsset) => {
if (!actorSource) return
setSubjectAssetBusy(`delete:${asset.id}`)
try {
const updated = await deleteSubjectAsset(job.id, actorSource.frame.index, actorSource.element.id, asset.id)
onJobUpdate(updated)
toast.success("主体视图已删除")
} catch (e) {
toast.error("主体视图删除失败:" + (e instanceof Error ? e.message : String(e)))
} finally {
setSubjectAssetBusy(null)
}
}
const updateFramePreviewPosition = (event: ReactMouseEvent<HTMLDivElement>, frameIndex: number) => { const updateFramePreviewPosition = (event: ReactMouseEvent<HTMLDivElement>, frameIndex: number) => {
const margin = 16 const margin = 16
const previewWidth = Math.min(340, window.innerWidth - margin * 2) const previewWidth = Math.min(340, window.innerWidth - margin * 2)
@@ -1635,7 +1695,7 @@ function SourceReferenceBuildPanel({
<span className="text-[10.5px] text-white/30"></span> <span className="text-[10.5px] text-white/30"></span>
</div> </div>
<div className="mt-2 grid grid-cols-6 gap-1.5 md:grid-cols-8 xl:grid-cols-12 2xl:grid-cols-16"> <div className="mt-2 grid grid-cols-[repeat(auto-fill,minmax(38px,1fr))] gap-1">
{frames.map((frame, index) => { {frames.map((frame, index) => {
const selected = selectedFrames.has(frame.index) const selected = selectedFrames.has(frame.index)
return ( return (
@@ -1715,7 +1775,7 @@ function SourceReferenceBuildPanel({
placeholder="统一方向:如年轻女性 / 更运动 / 更高级" placeholder="统一方向:如年轻女性 / 更运动 / 更高级"
className="h-7 w-[240px] min-w-[180px] rounded-md border border-white/10 bg-black/35 px-2 text-[10.5px] text-white outline-none placeholder:text-white/28 focus:border-cyan-300/50" className="h-7 w-[240px] min-w-[180px] rounded-md border border-white/10 bg-black/35 px-2 text-[10.5px] text-white outline-none placeholder:text-white/28 focus:border-cyan-300/50"
/> />
<span>{actorAssets.length}/6</span> <span>{visibleActorAssets.length}/6</span>
<button <button
type="button" type="button"
onClick={() => void generateSimilarActor()} onClick={() => void generateSimilarActor()}
@@ -1727,23 +1787,60 @@ function SourceReferenceBuildPanel({
</button> </button>
</div> </div>
</div> </div>
{actorAssets.length ? ( {visibleActorAssets.length ? (
<div className="grid grid-cols-6 gap-1.5"> <div className="flex flex-wrap gap-1.5">
{actorAssets.slice(-6).map((asset) => ( {visibleActorAssets.map((asset) => {
<a const busyMode = subjectAssetBusy?.endsWith(asset.id) ? subjectAssetBusy.split(":")[0] : ""
return (
<div
key={asset.id} key={asset.id}
href={subjectAssetUrl(job, asset)}
target="_blank"
rel="noreferrer"
onMouseEnter={(event) => updateSubjectAssetPreviewPosition(event, asset.id)} onMouseEnter={(event) => updateSubjectAssetPreviewPosition(event, asset.id)}
onMouseMove={(event) => updateSubjectAssetPreviewPosition(event, asset.id)} onMouseMove={(event) => updateSubjectAssetPreviewPosition(event, asset.id)}
onMouseLeave={() => setSubjectAssetPreview(null)} onMouseLeave={() => setSubjectAssetPreview(null)}
className="h-20 overflow-hidden rounded border border-white/10 bg-white transition hover:border-cyan-200/70 2xl:h-24" className="group relative aspect-[9/16] w-12 overflow-hidden rounded border border-white/10 bg-white transition hover:border-cyan-200/70 2xl:w-14"
title={asset.label || asset.view} title={asset.label || asset.view}
> >
<img src={subjectAssetUrl(job, asset)} alt={asset.label || asset.view} className="h-full w-full object-contain" /> <a
</a> href={subjectAssetUrl(job, asset)}
))} target="_blank"
rel="noreferrer"
className="absolute inset-0"
>
<img src={subjectAssetUrl(job, asset)} alt={asset.label || asset.view} className="h-full w-full object-contain" />
</a>
<div className="absolute bottom-1 right-1 flex flex-col gap-1 opacity-0 transition group-hover:opacity-100 group-focus-within:opacity-100">
<button
type="button"
onClick={(event) => {
event.preventDefault()
event.stopPropagation()
void regenerateSubjectAsset(asset)
}}
disabled={!!subjectAssetBusy}
className="inline-flex h-5 w-5 items-center justify-center rounded-full border border-cyan-100/35 bg-black/78 text-cyan-100 transition hover:border-cyan-100/70 hover:bg-cyan-500/25 focus:outline-none focus:ring-1 focus:ring-cyan-100/70 disabled:cursor-not-allowed disabled:opacity-60"
aria-label={`重新生成${asset.label || asset.view}`}
title="重新生成这一张"
>
{busyMode === "regen" ? <Loader2 className="h-3 w-3 animate-spin" /> : <RefreshCw className="h-3 w-3" />}
</button>
<button
type="button"
onClick={(event) => {
event.preventDefault()
event.stopPropagation()
void deleteActorAsset(asset)
}}
disabled={!!subjectAssetBusy}
className="inline-flex h-5 w-5 items-center justify-center rounded-full border border-rose-100/35 bg-black/78 text-rose-100 transition hover:border-rose-100/70 hover:bg-rose-500/25 focus:outline-none focus:ring-1 focus:ring-rose-100/70 disabled:cursor-not-allowed disabled:opacity-60"
aria-label={`删除${asset.label || asset.view}`}
title="删除这一张"
>
{busyMode === "delete" ? <Loader2 className="h-3 w-3 animate-spin" /> : <Trash2 className="h-3 w-3" />}
</button>
</div>
</div>
)
})}
</div> </div>
) : ( ) : (
<div className="rounded border border-dashed border-white/12 px-2 py-2 text-[10.5px] leading-snug text-white/32"> <div className="rounded border border-dashed border-white/12 px-2 py-2 text-[10.5px] leading-snug text-white/32">

View File

@@ -1066,6 +1066,7 @@ export async function generateSubjectAssets(
subject_style?: "transparent_human" | "source_actor" subject_style?: "transparent_human" | "source_actor"
reconstruction_mode?: "same" | "similar" reconstruction_mode?: "same" | "similar"
prompt?: string prompt?: string
replace_views?: boolean
} = {}, } = {},
): Promise<Job> { ): Promise<Job> {
const res = await fetch(`${API_BASE}/jobs/${jobId}/frames/${frameIdx}/elements/${elementId}/subject-assets`, { const res = await fetch(`${API_BASE}/jobs/${jobId}/frames/${frameIdx}/elements/${elementId}/subject-assets`, {
@@ -1081,6 +1082,7 @@ export async function generateSubjectAssets(
subject_style: body.subject_style ?? "transparent_human", subject_style: body.subject_style ?? "transparent_human",
reconstruction_mode: body.reconstruction_mode ?? "same", reconstruction_mode: body.reconstruction_mode ?? "same",
prompt: body.prompt ?? "", prompt: body.prompt ?? "",
replace_views: body.replace_views ?? false,
}), }),
}) })
if (!res.ok) { if (!res.ok) {
@@ -1089,3 +1091,19 @@ export async function generateSubjectAssets(
} }
return res.json() return res.json()
} }
export async function deleteSubjectAsset(
jobId: string,
frameIdx: number,
elementId: string,
assetId: string,
): Promise<Job> {
const res = await fetch(`${API_BASE}/jobs/${jobId}/frames/${frameIdx}/elements/${elementId}/subject-assets/${assetId}`, {
method: "DELETE",
})
if (!res.ok) {
const txt = await res.text().catch(() => "")
throw new Error(`deleteSubjectAsset ${res.status} ${txt.slice(0, 300)}`)
}
return res.json()
}