auto-save 2026-05-14 07:39 (~4)

api/main.py

@@ -2540,6 +2540,10 @@ class GenerateStoryboardVideoReq(BaseModel):
     size: str = "720x1280"
 
 
+class ProductFusionDescriptionReq(BaseModel):
+    shots: list[ProductFusionShot] = Field(default_factory=list)
+
+
 def video_seconds(duration: float) -> str:
     if video_uses_ark():
         if duration <= 0:
@@ -2997,6 +3001,59 @@ def create_product_fusion_guide(job_id: str, req: ProductFusionShot) -> dict:
     }
 
 
+def fallback_product_fusion_descriptions() -> list[str]:
+    return [
+        "人物双手拿起 SKG 颈部按摩仪，准备戴到脖子上，镜头轻微推近产品。",
+        "人物把 SKG 按摩仪贴合到肩颈位置，手部轻轻调整两侧机身角度。",
+        "人物坐在场景中轻按侧边控制区，产品保持在画框指定区域内清晰可见。",
+        "人物闭眼放松，肩颈从紧绷变舒展，产品佩戴位置稳定不漂移。",
+        "镜头靠近展示 SKG 产品材质、按键和内侧触点，手部不要遮挡产品主体。",
+        "使用后的放松状态收尾，人物自然抬头，产品仍保持白色 U 形外观和真实比例。",
+    ]
+
+
+@app.post("/jobs/{job_id}/product-fusion/descriptions")
+def generate_product_fusion_descriptions(job_id: str, req: ProductFusionDescriptionReq) -> dict:
+    if job_id not in JOBS:
+        raise HTTPException(404, "job not found")
+    fallback = fallback_product_fusion_descriptions()
+    shots = (req.shots or [])[:6]
+    if not LLM_API_KEY:
+        return {"descriptions": fallback, "mode": "fallback"}
+    shot_lines = []
+    for i, shot in enumerate(shots, start=1):
+        product = (shot.product_image or {}).get("label") or "SKG 产品图"
+        person = (shot.person_image or {}).get("label") or "白底人物姿态图"
+        scene = (shot.scene_image or {}).get("label") or "场景图"
+        region = shot.product_region
+        region_text = f"x={region.x:.2f}, y={region.y:.2f}, w={region.w:.2f}, h={region.h:.2f}" if region else "未画区域"
+        shot_lines.append(f"{i}. 产品={product}；人物={person}；区域={region_text}；场景={scene}；已有描述={shot.action_text or '空'}")
+    prompt = (
+        "你是 SKG 产品短视频分镜导演。请为 6 条产品融合镜头各写一条中文动作描述，"
+        "每条 20-40 字，必须说明人物在做什么、产品如何佩戴/展示、动作如何自然连续。"
+        "产品是 SKG 白色 U 形颈部/肩颈按摩仪，不要写医疗治疗承诺，不要出现竞品。"
+        "输出 JSON：{\"descriptions\":[\"...\", \"...\"]}。\n\n"
+        + "\n".join(shot_lines)
+    )
+    try:
+        resp = llm().chat.completions.create(
+            model=REWRITE_MODEL,
+            messages=[
+                {"role": "system", "content": "只输出合法 JSON，不要解释。"},
+                {"role": "user", "content": prompt},
+            ],
+            temperature=0.5,
+        )
+        text = resp.choices[0].message.content or ""
+        data = json.loads(text)
+        descriptions = [str(x).strip() for x in data.get("descriptions", []) if str(x).strip()]
+        if len(descriptions) < 6:
+            descriptions = (descriptions + fallback)[:6]
+        return {"descriptions": descriptions[:6], "mode": "llm"}
+    except Exception:
+        return {"descriptions": fallback, "mode": "fallback"}
+
+
 @app.get("/jobs/{job_id}/assets/{asset_id}.jpg")
 def get_storyboard_asset(job_id: str, asset_id: str):
     p = job_dir(job_id) / "assets" / f"{asset_id}.jpg"