feat: 优化模特上身图提示词

2026-07-14 15:18:54 +08:00
parent da58f99293
commit 0a58b89919
8 changed files with 1973 additions and 15 deletions
@@ -779,14 +779,29 @@ def _ratio_to_image_size(ratio: str) -> str:
    这里只能取最接近的「竖/横/方」近似;要拿到精确像素比例需在出图后做主体保护裁切/补边(见优化文档 §7.1
    normalize_output_image,属待部署的后处理项)。修复点:4:5 原会 fallback 成 1024x1024(方图)→ 比例错误,
    现归到竖图 1024x1536。"""
-    return {
+    known = {
        "1:1": "1024x1024",
        "3:4": "1024x1536",   # 近似竖图(网关无精确 3:4)
        "4:5": "1024x1536",   # 近似竖图(原 fallback 成方图是比例 bug)
        "9:16": "1024x1536",  # 近似竖图(网关无精确 9:16)
        "4:3": "1536x1024",
        "16:9": "1536x864",  # 真 16:9(原来误用 1536x1024 = 3:2)
-    }.get((ratio or "").strip(), "1024x1024")
+    }
+    normalized = (ratio or "").strip()
+    if normalized in known:
+        return known[normalized]
+    # 手动宽高比不能被静默当成方图。GPT 网关不接受任意精确尺寸时，至少保持横/竖方向；
+    # 原始比例仍写进有效提示词，供应商获得的是它支持的最近画布。
+    try:
+        width, height = (float(part.strip()) for part in normalized.split(":", 1))
+        if width > 0 and height > 0:
+            if width < height:
+                return "1024x1536"
+            if width > height:
+                return "1536x1024"
+    except (TypeError, ValueError):
+        pass
+    return "1024x1024"


 def _ratio_to_volcano_size(ratio: str) -> str:
@@ -990,6 +1005,93 @@ def build_model_tryon_prompt_refs(product, has_model: bool, base_prompt: str = "
    return " ".join(lines)


+def _model_tryon_prompt_v2_rollout_source(*, team_id, payload: dict) -> str | None:
+    """返回 V2.2 启用来源；未命中时失败关闭并继续使用旧提示词。"""
+
+    if payload.get("tryon_prompt_v2_override") is True:
+        return "internal_ab"
+    if getattr(settings, "MODEL_TRYON_PROMPT_V2_ENABLED", False):
+        return "global"
+
+    raw_allowlist = getattr(settings, "MODEL_TRYON_PROMPT_V2_CANARY_TEAM_IDS", ()) or ()
+    if isinstance(raw_allowlist, str):
+        raw_allowlist = raw_allowlist.split(",")
+    allowed_team_ids = {
+        str(value).strip().lower()
+        for value in raw_allowlist
+        if str(value).strip()
+    }
+    team_key = str(team_id or "").strip().lower()
+    return "canary" if team_key and team_key in allowed_team_ids else None
+
+
+def _build_model_tryon_prompt_v2(
+    *,
+    product,
+    payload: dict,
+    has_model: bool,
+    index: int,
+    n_product: int,
+    rollout_source: str,
+):
+    """使用任务创建时的分类快照构建单张有效提示词，并返回可追溯信息。"""
+
+    from apps.ai.tryon_prompt import (
+        ClassificationResult,
+        ProductContext,
+        build_tryon_prompt_plan,
+        classify_product_details,
+        default_ratio_for_kind,
+    )
+
+    batch_count = int(payload.get("tryon_batch_count") or 0)
+    if batch_count not in (1, 2, 4):
+        raise ValueError("unsupported_tryon_batch_count")
+    if index < 0 or index >= batch_count:
+        raise ValueError("tryon_index_out_of_range")
+
+    context = ProductContext.create(
+        title=getattr(product, "title", "") or "商品",
+        category=getattr(product, "category", "") or "",
+        description=getattr(product, "description", "") or "",
+        selling_points=tuple(product.selling_points.values_list("title", flat=True)[:8]),
+    )
+    classification = ClassificationResult.from_payload(payload.get("tryon_classification"))
+    if classification is None:
+        classification = classify_product_details(context, str(payload.get("prompt") or ""))
+
+    requested_ratio = str(payload.get("ratio") or "").strip()
+    resolved_ratio = requested_ratio or default_ratio_for_kind(classification.kind)
+    plan = build_tryon_prompt_plan(
+        context=context,
+        user_prompt=str(payload.get("prompt") or ""),
+        count=batch_count,
+        ratio=resolved_ratio,
+        product_reference_count=n_product,
+        has_model_portrait=has_model,
+        has_model_triview=False,
+        classification=classification,
+    )
+    prompt = plan.prompts[index]
+    trace = {
+        "version": "v2.2",
+        "applied": True,
+        "rollout_source": rollout_source,
+        "effective_prompt": prompt,
+        "shot_index": index,
+        "batch_count": batch_count,
+        "requested_ratio": requested_ratio or None,
+        "resolved_ratio": resolved_ratio,
+        "ratio_source": "user" if requested_ratio else "default",
+        "reference_roles": {
+            "product_numbers": list(plan.references.product_numbers),
+            "model_portrait_number": plan.references.model_portrait_number,
+            "model_triview_number": plan.references.model_triview_number,
+        },
+    }
+    return prompt, trace, resolved_ratio
+
+
 # 平台套图同批多张要「同款商品、不同版式」,否则 N 张文案/构图雷同(PMC#24)。
 # 锁死商品一致性,只让排版/构图/视角/配色基调按张变化。
 # 注:仅纯文生图回落路径(无参考图)仍用这个简表;refs 版改用下面的 slot 体系。
@@ -2649,7 +2751,7 @@ def _reap_stale_standalone_image_tasks(*, team) -> None:
            continue


-def enqueue_standalone_images(*, team, user, prompt: str, mode: str = "image", count: int = 1, product_id: str | None = None, reference_product: bool = False, model_id: str | None = None, model_entity_id: str | None = None, ratio: str | None = None, image_model: str | None = None, conversation=None, reference_image_ids: list[str] | None = None, platform_id: str | None = None, batch_id: str | None = None) -> list[AITask]:
+def enqueue_standalone_images(*, team, user, prompt: str, mode: str = "image", count: int = 1, product_id: str | None = None, reference_product: bool = False, model_id: str | None = None, model_entity_id: str | None = None, ratio: str | None = None, image_model: str | None = None, conversation=None, reference_image_ids: list[str] | None = None, platform_id: str | None = None, batch_id: str | None = None, tryon_prompt_v2_override: bool = False, tryon_ab: dict | None = None, dispatch: bool = True) -> list[AITask]:
    """独立生图(图片创作 / 模特上身图 / 平台套图)改为**异步**:本函数在 Web 请求里只做「建任务 +
    预留额度」这种秒级的活,真正 ~30s 的 ARK 出图交给 Celery worker(generate_standalone_image_task)。

@@ -2666,10 +2768,10 @@ def enqueue_standalone_images(*, team, user, prompt: str, mode: str = "image", c
    task_type = _STANDALONE_TASK_TYPE.get(mode, AITask.Type.PRODUCT_IMAGE)
    count = max(1, min(int(count or 1), 12))
    ref_ids = [str(r) for r in (reference_image_ids or []) if r]
-    # 带参考图的图片创作必须走 image_edit(gpt-image)才真的还原上传素材——火山 Seedream 的图生图只松散
-    # 借色调、不锁主体(与三视图/模特上身图同理,那两条流程也强制 image_edit)。当前模型不支持 image_edit
-    # 就自动切到 gpt-image-2,否则用户选了火山却传了参考图 → 生成完全不参考素材(用户实测的根因)。
-    if ref_ids and not hasattr(build_provider(model_config), "image_edit"):
+    # 普通图片创作带用户上传参考图时，当前模型不支持 image_edit 就沿用既有保护切到 gpt-image。
+    # 模特上身图例外：商品图和模特图由专用 Worker 作为多参考图传入 Seedream/GPT，必须尊重用户自选模型；
+    # 即使旧客户端额外带了 reference_image_ids，也不得在任务创建阶段静默覆盖 image_model。
+    if ref_ids and mode != "model" and not hasattr(build_provider(model_config), "image_edit"):
        alt = resolve_image_model("gpt-image")
        if alt is not None:
            model_config = alt
@@ -2691,10 +2793,49 @@ def enqueue_standalone_images(*, team, user, prompt: str, mode: str = "image", c
    platform_name = _PLATFORM_NAMES.get(platform_key, "")
    from apps.billing.pricing import quote_flat

+    # Step 2.1:只为“模特上身图 + 商品”记录一次确定性分类快照。这里不读取图片、不调用模型；
+    # Step 3 的 Worker 根据 MODEL_TRYON_PROMPT_V2_ENABLED 决定使用 V2 或旧提示词。
+    tryon_classification: dict[str, str | None] | None = None
+    if mode == "model" and product_id:
+        from apps.ai.tryon_prompt import ClassificationResult, ProductContext, ProductKind, classify_product_details
+        from apps.products.models import Product
+
+        product_for_classification = Product.objects.filter(id=product_id, team=team).only(
+            "title", "category", "description"
+        ).first()
+        if product_for_classification is None:
+            classification = ClassificationResult(
+                ProductKind.UNKNOWN,
+                "fallback",
+                None,
+                "product_not_found",
+            )
+        else:
+            classification = classify_product_details(
+                ProductContext.create(
+                    title=product_for_classification.title,
+                    category=product_for_classification.category,
+                    description=product_for_classification.description,
+                ),
+                prompt,
+            )
+        tryon_classification = classification.as_payload()
+
    tasks: list[AITask] = []
    for index in range(count):
        quote = quote_flat(model_config, team=team)
        request_payload = {"model": model_config.name, "endpoint": model_config.endpoint, "prompt": prompt, "mode": mode, "index": index, "product_id": str(product_id) if product_id else None, "reference_product": bool(reference_product), "model_id": str(model_id) if model_id else None, "model_entity_id": str(model_entity_id) if model_entity_id else None, "batch_id": batch_id, "ratio": str(ratio) if ratio else None, "reference_image_ids": ref_ids, "platform_id": platform_key or None, "platform_name": platform_name or None}
+        if tryon_classification is not None:
+            request_payload["tryon_classification"] = dict(tryon_classification)
+            request_payload["tryon_batch_count"] = count
+            # 仅供内部 A/B 工具使用；GenerateImageView 不接收这两个参数，用户请求无法绕过全局开关。
+            if tryon_prompt_v2_override:
+                request_payload["tryon_prompt_v2_override"] = True
+            if tryon_ab:
+                request_payload["tryon_ab"] = {
+                    "experiment_id": str(tryon_ab.get("experiment_id") or ""),
+                    "variant": str(tryon_ab.get("variant") or ""),
+                }
        # 只在重跑/补图时落键(不落 False):workbench 用 KeyTextTransform 抽文本,"false" 字符串也是真值,会误判
        if is_append:
            request_payload["batch_append"] = True
@@ -2719,8 +2860,9 @@ def enqueue_standalone_images(*, team, user, prompt: str, mode: str = "image", c
        task.save(update_fields=["status", "updated_at"])
        tasks.append(task)
    # 额度都预留成功后再统一派发,避免"派发了任务但后面某张预留失败"的半成品状态
-    for task in tasks:
-        generate_standalone_image_task.delay(str(task.id))
+    if dispatch:
+        for task in tasks:
+            generate_standalone_image_task.delay(str(task.id))
    return tasks


@@ -2732,7 +2874,7 @@ def run_standalone_image_task(*, task_id: str) -> None:
        return
    team = task.team
    user = task.created_by
-    payload = task.request_payload or {}
+    payload = dict(task.request_payload or {})
    prompt = str(payload.get("prompt") or "")
    mode = str(payload.get("mode") or "image")
    index = int(payload.get("index") or 0)
@@ -2755,7 +2897,7 @@ def run_standalone_image_task(*, task_id: str) -> None:
    if product_id:
        from apps.products.models import Product

-        product = Product.objects.filter(id=product_id).first()
+        product = Product.objects.filter(id=product_id, team=team).first()
    can_edit = hasattr(provider, "image_edit")
    model_url = ""
    if payload.get("model_id"):
@@ -2777,13 +2919,49 @@ def run_standalone_image_task(*, task_id: str) -> None:
    # 参考图收集与 provider 无关:先把「该用哪些参考图 + 哪条提示词」定下来,再按模型能力选调用方式。
    edit_images: list[str] = []
    edit_prompt = ""
+    output_ratio = str(payload.get("ratio") or "")
+    tryon_prompt_trace: dict | None = None
+    tryon_rollout_source = (
+        _model_tryon_prompt_v2_rollout_source(team_id=team.id, payload=payload)
+        if mode == "model"
+        else None
+    )
    if mode == "model" and product_urls:
        # 模特上身图:参考图1~N=商品真实图(多角度),参考图N+1=模特(模特图可缺则让模型自取真人模特)
        edit_images = product_urls + ([model_url] if model_url else [])
-        edit_prompt = build_model_tryon_prompt_refs(
+        legacy_prompt = build_model_tryon_prompt_refs(
            product, has_model=bool(model_url), base_prompt=prompt,
            index=index, n_product=len(product_urls),
        )
+        edit_prompt = legacy_prompt
+        if tryon_rollout_source is not None:
+            try:
+                edit_prompt, tryon_prompt_trace, output_ratio = _build_model_tryon_prompt_v2(
+                    product=product,
+                    payload=payload,
+                    has_model=bool(model_url),
+                    index=index,
+                    n_product=len(product_urls),
+                    rollout_source=tryon_rollout_source,
+                )
+            except Exception as exc:  # noqa: BLE001 — 规划器异常时安全回落旧提示词，不让已预扣任务悬空
+                tryon_prompt_trace = {
+                    "version": "v2.2",
+                    "applied": False,
+                    "rollout_source": tryon_rollout_source,
+                    "effective_prompt": legacy_prompt,
+                    "shot_index": index,
+                    "batch_count": payload.get("tryon_batch_count"),
+                    "requested_ratio": output_ratio or None,
+                    "resolved_ratio": output_ratio or None,
+                    "ratio_source": "user" if output_ratio else None,
+                    "fallback_reason": str(exc) if isinstance(exc, ValueError) else type(exc).__name__,
+                    "reference_roles": {
+                        "product_numbers": list(range(1, len(product_urls) + 1)),
+                        "model_portrait_number": len(product_urls) + 1 if model_url else None,
+                        "model_triview_number": None,
+                    },
+                }
    elif mode == "cover" and product_urls:
        # 平台套图:参考图1~N=商品真实图(多角度,_product_reference_urls 已优先真实上传图/排除 AI 图),
        # 有模特则参考图N+1=模特(锁人脸/身形)。platform_id 注入平台版式块(优化版)。
@@ -2804,19 +2982,42 @@ def run_standalone_image_task(*, task_id: str) -> None:
        # 否则图生图只会松散借个色调、不真的还原上传的素材(=用户反馈的「没参考我的图」)。
        edit_images = ref_urls
        edit_prompt = build_free_reference_prompt(prompt, n_refs=len(ref_urls), index=index)
+    elif mode == "model" and product_id and tryon_rollout_source is not None:
+        # 没有真实商品参考图时维持原纯文生图行为，但明确留下未应用原因，不能伪装成 V2 已生效。
+        tryon_prompt_trace = {
+            "version": "v2.2",
+            "applied": False,
+            "rollout_source": tryon_rollout_source,
+            "effective_prompt": prompt,
+            "shot_index": index,
+            "batch_count": payload.get("tryon_batch_count"),
+            "requested_ratio": output_ratio or None,
+            "resolved_ratio": output_ratio or None,
+            "ratio_source": "user" if output_ratio else None,
+            "fallback_reason": "no_product_reference",
+            "reference_roles": {
+                "product_numbers": [],
+                "model_portrait_number": None,
+                "model_triview_number": None,
+            },
+        }
    use_edit = bool(edit_images)
    try:
+        if tryon_prompt_trace is not None:
+            payload["tryon_prompt"] = tryon_prompt_trace
+            task.request_payload = payload
+            task.save(update_fields=["request_payload", "updated_at"])
        if use_edit and can_edit:
            # gpt-image 等支持 image_edit(多图参考编辑接口)
            if payload.get("reference_product"):
                size = "1536x1024"  # 三视图固定横向
            else:
-                size = _ratio_to_image_size(str(payload.get("ratio") or ""))  # 模特图按选中比例
+                size = _ratio_to_image_size(output_ratio)  # 模特图按用户比例或 V2 默认比例
            response = provider.image_edit(model=model_config.name, prompt=edit_prompt, images=edit_images, size=size)
        elif use_edit:
            # 火山 Seedream 无 image_edit:走 image_generation 带 image=参考图(图生图多参考);
            # 尺寸按选中比例换算成火山可接受的 ~2K 尺寸(三视图固定横向)
-            vsize = "2304x1728" if payload.get("reference_product") else _ratio_to_volcano_size(str(payload.get("ratio") or ""))
+            vsize = "2304x1728" if payload.get("reference_product") else _ratio_to_volcano_size(output_ratio)
            response = provider.image_generation(model=model_config.name, endpoint=model_config.endpoint, prompt=edit_prompt, image=edit_images, size=vsize)
        else:
            # 纯文生图同批多张要不同构图,否则雷同(PMC#24);index>0 追加换版式指令。