fix(#2187): outfit提取正则优化，保留颜色形容词 (#2187)

Co-authored-by: xiaoxia <dev@xiaoxiajianji.com>
Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>

apps/worker/worker_app/tasks/viral_video.py

@@ -404,6 +404,190 @@ def _analyze_single_image(
         product_nodes = [n for n in nodes if n["tag"] == "product"]
         scene = xp.text_of(raw_text, "scene") or "通用"
         mood = xp.text_of(raw_text, "mood") or ""
+        # #2184: #2177 XML 重构后人物信息放在顶层 <people has_person count gender age_range pose expression/>，
+        # 不再是 <product> 的 portrait_prompt 属性。需从顶层 people 标签提取并拼装 portrait_prompt。
+        portrait_prompt = "无人像"
+        try:
+            people_node = xp.find_first(raw_text, "people")
+            if people_node:
+                _pa = people_node.get("attrs") or {}
+                _has_person = xp.attr_bool(_pa.get("has_person"), False)
+                if _has_person:
+                    _gender = _pa.get("gender", "无法判断") or "无法判断"
+                    _age = _pa.get("age_range", "无法判断") or "无法判断"
+                    _hair = _pa.get("hair", "无法判断") or "无法判断"
+                    _skin = _pa.get("skin_tone", "无法判断") or "无法判断"
+                    _face = _pa.get("face_shape", "无法判断") or "无法判断"
+                    _outfit = _pa.get("outfit", "无法判断") or "无法判断"
+                    _pose = _pa.get("pose", "无法判断") or "无法判断"
+                    _expr = _pa.get("expression", "无法判断") or "无法判断"
+                    _count = xp.attr_int(_pa.get("count"), 1)
+                    # #2185: VLM有时对外貌属性输出"无法判断"，用通用兜底值确保portrait_prompt始终有完整外貌描述
+                    if _hair == "无法判断":
+                        _hair = "自然发型"
+                    if _skin == "无法判断":
+                        _skin = "自然"
+                    if _face == "无法判断":
+                        _face = "标准"
+                    if _outfit == "无法判断":
+                        _outfit = "日常服装"
+                    _parts = []
+                    if _gender != "无法判断":
+                        _g = _gender + ("性" if not _gender.endswith("性") else "")
+                        _parts.append(_g)
+                    else:
+                        _parts.append("成年人")
+                    if _age != "无法判断":
+                        _parts.append(_age)
+                    _parts.append("人物")
+                    _parts.append(_hair)
+                    _parts.append(f"{_skin}肤色")
+                    _parts.append(f"{_face}脸型")
+                    _parts.append(f"身着{_outfit}")
+                    if _pose != "无法判断":
+                        _parts.append(f"姿态{_pose}")
+                    if _expr != "无法判断":
+                        _parts.append(f"表情{_expr}")
+                    else:
+                        _parts.append("表情自然")
+                    # #2186: 智能回填——VLM有时省略hair/outfit等外貌属性，但product.name/features/colors里已有相关信息
+                    # 从product名字和features中提取服装关键词回填outfit
+                    if _outfit in ("日常服装", "无法判断"):
+                        for _ppn in product_nodes:
+                            _pn = (_ppn.get("attrs") or {}).get("name", "") or ""
+                            _pf = (_ppn.get("attrs") or {}).get("features", "") or ""
+                            _ptxt = _pn + " " + _pf
+                            # 服装关键词识别（常见上装/下装/裙装/套装）
+                            _cloth_kws = [
+                                "衬衫",
+                                "T恤",
+                                "毛衣",
+                                "针织衫",
+                                "卫衣",
+                                "外套",
+                                "西装",
+                                "夹克",
+                                "风衣",
+                                "大衣",
+                                "羽绒服",
+                                "马甲",
+                                "背心",
+                                "连衣裙",
+                                "半身裙",
+                                "短裙",
+                                "长裙",
+                                "牛仔裤",
+                                "休闲裤",
+                                "西裤",
+                                "运动裤",
+                                "短裤",
+                                "旗袍",
+                                "汉服",
+                                "制服",
+                                "polo衫",
+                                "POLO衫",
+                                "针织",
+                                "毛衫",
+                                "开衫",
+                                "帽衫",
+                                "皮夹克",
+                                "皮衣",
+                            ]
+                            for _ckw in _cloth_kws:
+                                if _ckw in _ptxt:
+                                    _ci = _ptxt.find(_ckw)
+                                    # 向前找颜色/材质/款式形容词（白/黑/米/红/蓝/灰/棉/麻/长/短/厚/薄/长袖/短袖/翻领/圆领/V领/印花/条纹等）
+                                    _start = max(0, _ci - 8)
+                                    # 向后包含款式词（长袖/短袖/外套/套装/上衣等后续修饰）
+                                    _end = min(len(_ptxt), _ci + len(_ckw) + 4)
+                                    _outfit_extract = _ptxt[_start:_end].strip(" ,，。.、")
+                                    # 仅清理明确的品牌/产品类前缀（不清理颜色/款式/尺寸形容词）
+                                    _outfit_extract = re.sub(
+                                        r"^(\S{0,4}牌|\S{0,3}品牌|\S{0,3}款|产品|商品|的)", "", _outfit_extract
+                                    ).strip()
+                                    # 尾部清理：去掉残留的品牌字/型号字（如"标""ml""g""装"等单字杂字）
+                                    _outfit_extract = re.sub(
+                                        r"(标[0-9a-zA-Z]*|\d+\s*(?:ml|g|L|斤|件|个|瓶|盒|包|袋|装)|\s+\d+\s*)$",
+                                        "",
+                                        _outfit_extract,
+                                        flags=re.IGNORECASE,
+                                    ).strip()
+                                    if len(_outfit_extract) >= 2:
+                                        _outfit = _outfit_extract
+                                        break
+                            if _outfit not in ("日常服装", "无法判断"):
+                                break
+                    # 从color标签中提取头发颜色回填hair
+                    if _hair in ("自然发型", "无法判断"):
+                        _hair_color = ""
+                        _color_nodes = [n for n in nodes if n["tag"] == "color"]
+                        _hair_kws_map = {
+                            "黑": "黑色",
+                            "棕": "棕色",
+                            "金": "金色",
+                            "栗": "栗色",
+                            "红": "红色",
+                            "白": "白色",
+                            "灰": "灰色",
+                            "蓝": "蓝色",
+                            "黄": "黄色",
+                            "紫": "紫色",
+                        }
+                        for _cn in _color_nodes:
+                            _cname = (_cn.get("attrs") or {}).get("name", "") or ""
+                            # 小占比颜色更可能是发色（非主色的小面积色），且名称含头发/黑/棕/金等
+                            _ccov = 0.0
+                            try:
+                                _ccov = float((_cn.get("attrs") or {}).get("coverage", "0") or 0)
+                            except Exception:
+                                pass
+                            for _hk, _hv in _hair_kws_map.items():
+                                if _hk in _cname and _ccov < 0.3:
+                                    _hair_color = _hv
+                                    break
+                            if _hair_color:
+                                break
+                        if _hair_color:
+                            _hair = f"{_hair_color}头发"
+                        else:
+                            _hair = "自然发型"
+                    # 重新拼装_parts（回填后）
+                    _parts = []
+                    if _gender != "无法判断":
+                        _g = _gender + ("性" if not _gender.endswith("性") else "")
+                        _parts.append(_g)
+                    else:
+                        _parts.append("成年人")
+                    if _age != "无法判断":
+                        _parts.append(_age)
+                    _parts.append("人物")
+                    _parts.append(_hair)
+                    _parts.append(f"{_skin}肤色")
+                    _parts.append(f"{_face}脸型")
+                    _parts.append(f"身着{_outfit}")
+                    if _pose != "无法判断":
+                        _parts.append(f"姿态{_pose}")
+                    if _expr != "无法判断":
+                        _parts.append(f"表情{_expr}")
+                    else:
+                        _parts.append("表情自然")
+                    portrait_prompt = "，".join(_parts)
+                    logger.info(
+                        "[爆款视频] 图片 #%d 解析<people>(回填后): count=%d gender=%s age=%s hair=%s skin=%s face=%s outfit=%s pose=%s expr=%s → %s",
+                        idx,
+                        _count,
+                        _gender,
+                        _age,
+                        _hair,
+                        _skin,
+                        _face,
+                        _outfit,
+                        _pose,
+                        _expr,
+                        portrait_prompt,
+                    )
+        except Exception as _pe:
+            logger.warning("[爆款视频] 图片 #%d 解析<people>标签异常: %s，回退无人像", idx, _pe)
         for p in product_nodes:
             a = p["attrs"]
             text_on_pkg = a.get("text_on_package", "")
@@ -419,6 +603,10 @@ def _analyze_single_image(
             appearance = a.get("appearance", "") or "无法判断"
             packaging = a.get("packaging", "") or "无法判断"
             summary = a.get("summary", "") or f"{brand} {name}"
+            # 优先取 product 属性上的 portrait_prompt（兼容旧schema），否则用顶层 <people> 解析结果
+            _pp_from_attr = a.get("portrait_prompt", "")
+            if _pp_from_attr and _pp_from_attr != "无人像":
+                portrait_prompt = _pp_from_attr
             return {
                 "name": name,
                 "brand": brand,
@@ -429,10 +617,26 @@ def _analyze_single_image(
                 "key_features": feat_list or [features] if features else ["无法判断"],
                 "scene": scene,
                 "mood": mood,
-                "portrait_prompt": a.get("portrait_prompt", "无人像"),
+                "portrait_prompt": portrait_prompt,
                 "summary": summary,
                 "_source": "xml",
             }
+        # 没有 product 标签但有 <people has_person="true"> 也要能取到人物描述（兜底）
+        if portrait_prompt != "无人像":
+            return {
+                "name": "未识别",
+                "brand": "无法判断",
+                "category": "无法判断",
+                "appearance": "无法判断",
+                "packaging": "无法判断",
+                "text_on_package": [],
+                "key_features": ["无法判断"],
+                "scene": scene,
+                "mood": mood,
+                "portrait_prompt": portrait_prompt,
+                "summary": "未识别",
+                "_source": "xml_no_product",
+            }
         return _vision_fallback(idx, "no_product_tag")
 
     def _normalize(raw, source: str) -> dict:
@@ -617,7 +821,7 @@ def _step_intent_parsing(job: ViralVideoJob, image_analysis: dict) -> dict:
                 temperature=0.4,
                 max_tokens=1024,
                 model=_m,
-                timeout=45,
+                timeout=60,
             )  # #2180: 意图解析 LLM 实测需更长响应，原25s太紧
             if not raw:
                 continue
@@ -1037,15 +1241,16 @@ def _step_script_generation(job: ViralVideoJob, intent: dict, image_analysis: di
     _pro = getattr(_s, "doubao_model", None) or _fast
     try:
         # 第一次：快模型 25s
-        normalized = _try_gen(_fast, 0.8, 2500, "fast-first", tmo=45)
+        normalized = _try_gen(_fast, 0.8, 2500, "fast-first", tmo=90)
         if normalized is not None:
             return normalized
-        # 第二次：快模型降温度+加大 max_tokens，25s
-        normalized = _try_gen(_fast, 0.6, 3200, "fast-retry", tmo=45)
+        # #2183: 实测pro 1500tok输出需75.8s，单次timeout提到90s
+        normalized = _try_gen(_fast, 0.6, 3200, "fast-retry", tmo=90)
         if normalized is not None:
             return normalized
+        # 第三次：用主力模型兜底，给 120s
         if _pro and _pro != _fast:
-            normalized = _try_gen(_pro, 0.7, 3500, "pro-fallback", tmo=60)
+            normalized = _try_gen(_pro, 0.7, 3500, "pro-fallback", tmo=120)
             if normalized is not None:
                 return normalized
         logger.warning("[爆款视频] 编导脚本三次都未生成合格结果，使用兜底脚本")
@@ -1310,13 +1515,47 @@ def _step_render(job: ViralVideoJob, copy_result: dict, tts_audio_url: str | Non
         pre_trusted = list(pti)
         logger.info("[爆款视频] 使用信任链预热结果 n=%d，跳过现场 Seedream AI 化", len(pre_trusted))
     elif all_portrait_urls and _mcfg.get("provider", "doubao") == "doubao":
+        # #2183: 真·现场跑信任链——同步调用 Seedream t2i，拿到 AI 人像 URL 后再传 Seedance
         logger.info(
-            "[爆款视频] 预热结果不可用(%s/%d张)，将现场跑信任链",
+            "[爆款视频] 预热结果不可用(%s/%d张)，现场同步跑信任链Seedream t2i",
             "缺失" if not pti else f"{len(pti)}/{len(all_portrait_urls)}",
             len(all_portrait_urls),
         )
+        try:
+            from packages.shared.ai_service import preheat_trust_chain
 
-    # 第一次调用：带参考图/首帧/音频/参考视频
+            _ia = getattr(job, "image_analysis", None) or {}
+            _prods = (_ia.get("products") if isinstance(_ia, dict) else None) or []
+            _pdescs = []
+            if _prods:
+                _pdescs = [(pp.get("portrait_prompt") or "无人像") for pp in _prods]
+            elif isinstance(_ia, dict):
+                _pp0 = _ia.get("portrait_prompt") or "无人像"
+                if _pp0 and _pp0 != "无人像":
+                    _pdescs = [_pp0]
+            _valid = [d for d in _pdescs if d and isinstance(d, str) and "无人像" not in d and len(d) >= 10]
+            if _valid:
+                _t0 = time.time()
+                _live_urls = preheat_trust_chain(_valid, timeout=120)
+                if _live_urls and len(_live_urls) == len(all_portrait_urls):
+                    pre_trusted = list(_live_urls)
+                    logger.info(
+                        "[爆款视频] 现场信任链t2i完成 %d张 耗时%.1fs，将用AI人像传Seedance",
+                        len(pre_trusted),
+                        time.time() - _t0,
+                    )
+                else:
+                    logger.warning(
+                        "[爆款视频] 现场信任链t2i返回不匹配 urls=%s n_portraits=%d，回退原图+400降级纯t2v",
+                        _live_urls,
+                        len(all_portrait_urls),
+                    )
+            else:
+                logger.info("[爆款视频] 无有效人物描述（可能是商品图），无需现场跑信任链")
+        except Exception as _te:
+            logger.warning("[爆款视频] 现场跑信任链异常: %s，回退原图+400降级纯t2v", _te, exc_info=True)
+
+    # 第一次调用：带参考图/首帧/音频/参考视频（pre_trusted有值→走信任链；无值→原图；若400 ai_client内部自动降级纯t2v）
     result = call_video_generation(
         prompt=prompt,
         image_url=first_image,
@@ -1609,8 +1848,26 @@ def run_viral_video_analyze(self: Task, job_id: str) -> dict:
 
         image_analysis = _step_image_analysis(job)
         job.image_analysis = image_analysis
+        # #2183/#2174: VLM完成后立即启动信任链t2i预热（后台daemon线程，与后续阶段并行）
+        if job.images:
+            try:
+                _products = (image_analysis or {}).get("products", []) or []
+                _portrait_descs = [(p.get("portrait_prompt") or "无人像") for p in _products] if _products else []
+                if not _portrait_descs and isinstance(image_analysis, dict):
+                    _pp = image_analysis.get("portrait_prompt") or "无人像"
+                    if _pp and _pp != "无人像":
+                        _portrait_descs = [_pp]
+                _start_trust_chain_preheat(job.id, _portrait_descs)
+            except Exception as _e:
+                logger.warning("[爆款视频][阶段1] 启动信任链t2i预热失败: %s", _e)
         _save_job(repo, job, session)
-        _emit_progress(job_id, ViralVideoStage.IMAGE_ANALYSIS, 60.0, "图片分析完成", {"result": image_analysis})
+        _emit_progress(
+            job_id,
+            ViralVideoStage.IMAGE_ANALYSIS,
+            60.0,
+            "图片分析完成，AI人像预热中...",
+            {"result": image_analysis, "trust_chain_preheating": True},
+        )
 
         style_guide = None
         if job.reference_video_url or job.style_template_id:
@@ -1657,8 +1914,8 @@ def run_viral_video_analyze(self: Task, job_id: str) -> dict:
     bind=True,
     max_retries=1,
     name="worker.run_viral_video_generate_copy",
-    soft_time_limit=360,  # #2173: 6min（编导脚本含意图+三级重试+审核，fast超时转pro）
-    time_limit=420,  # #2173: 7min hard limit
+    soft_time_limit=600,  # #2183: pro长脚本~120s+三级重试最坏300s+intent/review ~105s，给到10min
+    time_limit=660,  # #2183: 11min hard limit
 )
 def run_viral_video_generate_copy(self: Task, job_id: str) -> dict:
     """v1.6 阶段2（v1.6.1 提速版）：意图解析 → 编导分镜脚本生成 → 直接返回，合规审核后置到出片前。

packages/application/viral_video/prompts.py

@@ -50,14 +50,25 @@ _IMAGE_ANALYSIS_SYSTEM = f"""你是电商商品视觉分析师，负责从商品
 请严格按下面的标签格式输出，标签名一个都不能改，不要输出任何解释，不要用代码块：
 <products> 下面每个产品用一个 <product> 标签，属性 name 是产品名、features 是外观特征、position 是 main 或 secondary、image_index 是第几张图（从0开始）。
 <colors> 下面每个主要颜色用一个 <color> 标签，属性 hex 是色值、name 是颜色名、coverage 是占比小数。
-<people> 用一个标签，属性 has_person、count、gender、age_range、pose、expression 分别描述人物情况。
+<people> 用一个标签，属性 has_person、count、gender、age_range、hair（发型发色）、skin_tone（肤色）、face_shape（脸型）、outfit（穿着）、pose（姿态）、expression（表情）分别描述人物外貌。有人物时属性尽量具体（如hair="黑色长直发"、outfit="白色衬衫"），无人像时除has_person=false外其他填"无法判断"。
 <mood> 标签写画面整体情绪氛围。
 <visible_text> 下面每处可见文字用一个 <text_item> 标签，属性 text 是文字内容、position 是位置。
 <scene> 标签写场景描述。
 <quality> 用一个标签，属性 resolution、lighting、composition、blur 描述画质。
 <key_selling_points> 下面每个卖点用一个 <point> 标签。
 
-看不到或无法判断的内容，属性值填“无法判断”，布尔值填 false，不要留空标签。"""
+【人物属性硬性要求（has_person=true时必须遵守）】
+hair/skin_tone/face_shape/outfit四项绝对禁止填“无法判断”，必须基于图片可见特征给出具体中文描述：
+- hair：必须描述发型+发色，如“黑色齐肩直发”“棕色微卷中长发”“深棕色短发”
+- skin_tone：必须描述肤色，如“暖调自然肤色”“白皙肤色”“小麦色”
+- face_shape：必须描述脸型，如“鹅蛋脸”“圆脸”“瓜子脸”“方脸”
+- outfit：必须描述可见穿着，如“米色翻领衬衫”“白色T恤”“黑色连衣裙”
+即使局部被遮挡也要根据可见部分合理推断；确实看不清时按最接近的直观印象描述。
+
+其他非人物属性看不到或无法判断时填“无法判断”，布尔值填false，不要留空标签。
+
+【有人物场景输出参考（女性手持商品示例，必须写全10个属性，禁止省略）】
+<people has_person="true" count="1" gender="女" age_range="青年" hair="黑色齐肩直发" skin_tone="暖调自然肤色" face_shape="鹅蛋脸" outfit="米色翻领衬衫" pose="正面半身，手持商品" expression="面带微笑"/>"""
 
 _IMAGE_ANALYSIS_USER = """请分析以下商品图片，共 {image_count} 张。
 所属行业：{industry}
@@ -73,7 +84,7 @@ _IMAGE_ANALYSIS_EXAMPLE = """<products>
 <color hex="#D32F2F" name="红色" coverage="0.4"/>
 <color hex="#FFFFFF" name="白色" coverage="0.5"/>
 </colors>
-<people has_person="false" count="0" gender="无法判断" age_range="无法判断" pose="无法判断" expression="无法判断"/>
+<people has_person="false" count="0" gender="无法判断" age_range="无法判断" hair="无法判断" skin_tone="无法判断" face_shape="无法判断" outfit="无法判断" pose="无法判断" expression="无法判断"/>
 <mood>干净、实用</mood>
 <visible_text>
 <text_item text="多功能油污净" position="瓶身正面"/>

packages/config/base.py

@@ -92,7 +92,7 @@ class SharedSettings(BaseSettings):
     doubao_api_key: str = ""
     doubao_model: str = "doubao-seed-2-1-pro-260915"  # 推理模型（Seed 2.1 Pro，深度思考+多模态；原 seed-1-6 已下线）
     doubao_fast_model: str = (
-        "doubao-seed-2-1-lite-260915"  # 快速模型（Seed 2.1 Lite，高 RPM，编导/审核/VLM lite；原 1-5-pro-32k 已 Retiring）
+        "doubao-seed-2-1-pro-260915"  # #2181: lite方舟侧100%超时，默认fast_model也走pro；方舟恢复lite后通过ENV DOUBAO_FAST_MODEL切回
     )
     doubao_base_url: str = "https://ark.cn-beijing.volces.com/api/v3"
     doubao_timeout: int = 45  # #2180: 方舟LLM高峰期响应6-8s，原30s太紧提到45s

packages/shared/ai_client.py

@@ -585,7 +585,7 @@ class DoubaoClient:
         # #2172/#2174: 信任链——使用预热好的 Seedream t2i 文生图（纯模型生成人像，是方舟信任产物，
         # 不会触发肖像审核）。预热在 VLM 分析后由 daemon 线程后台完成，结果通过 pre_trusted_images 传入。
         # - 预热结果有效 → 替换原参考图，走 omni_ref 模式
-        # - 预热结果不可用 → 直接用原图（若被400肖像拦截，#2166自动降级纯t2v），避免现场跑t2i阻塞渲染
+        # - 预热结果不可用 → 直接用原图（上层 _step_render 已现场同步跑 Seedream t2i 兜底；若再被400拦截，下方自动降级纯t2v）
         # 信任链只作用于 doubao provider；DashScope(Wan) 保持原行为。
         trust_chain_applied = False
         if provider == "doubao" and getattr(self, "trust_chain_enabled", True) and pre_trusted_images:
@@ -763,6 +763,26 @@ class DoubaoClient:
                 # 保留第二次的错误信息
                 sc, body = sc2, body2
 
+        # #2183: portrait_intercept / 真人肖像审核拦截 → 去掉所有参考图（含image_url首帧），纯 t2v 重试一次
+        # 信任链预热或现场 t2i 都失败时的最后兜底，保证能出片
+        if not task_id and sc == 400:
+            _err_code_for_400, _ = _classify_video_error(sc, body, last_err)
+            if _err_code_for_400 == "portrait_intercept" and (image_url or ref_imgs):
+                logger.warning(
+                    "Seedance 创建因 portrait_intercept 失败，降级纯 t2v（移除所有参考图）重试: img=%d ref=%d",
+                    1 if image_url else 0,
+                    len(ref_imgs),
+                )
+                _t2v_payload = dict(create_payload)
+                _t2v_payload["content"] = [{"type": "text", "text": prompt.strip()}]
+                _t2v_payload["ratio"] = ratio or "9:16"
+                task_id, last_err, sc3, body3 = _do_create(_t2v_payload)
+                if task_id:
+                    sc, body = sc3, body3
+                    logger.info("[trust-chain] portrait_intercept 降级纯 t2v 成功 task_id=%s", task_id)
+                else:
+                    sc, body = sc3, body3
+
         if not task_id:
             err_code, user_msg = _classify_video_error(sc, body, last_err)
             self.last_video_error = {