feat: AI生图功能丰富

server/services/generator.py

@@ -81,19 +81,21 @@ DEFAULT_REF_COUNT = 2  # 每次生图最多带的参考图数（正面 1 张 +
 def _image_size(provider: str, ratio: str, is_wan: bool = True, model: str = "") -> str:
     """平台比例 → provider 尺寸参数。3:4 竖版（Ozon），1:1 方图（国内）。"""
     if provider == "doubao":
-        return "1536x2048" if ratio == "3:4" else "2048x2048"
+        return {"3:4": "1536x2048", "4:5": "1536x1920"}.get(ratio, "2048x2048")
     if provider == "rightapi":
         # gpt-image 自定义尺寸约束：16 的倍数、长短边比 ≤ 3:1（1536x2048 合法）
-        return "1536x2048" if ratio == "3:4" else "2048x2048"
+        return {"3:4": "1536x2048", "4:5": "1536x1920"}.get(ratio, "2048x2048")
     if provider == "pcore":
         # 竖图必须与 Ozon 的 3:4 规格一致。1024x1536 实际是 2:3，会使成图明显偏瘦高。
-        return "1152x1536" if ratio == "3:4" else "1024x1024"
+        return {"3:4": "1152x1536", "4:5": "1280x1600"}.get(ratio, "1024x1024")
     # tongyi：万象与千问的 size 语法相同（* 分隔），档位不同
     # wan2.6 系列总像素限制在 [1280², 1440²]，wan2.7 的 1536*2048/2048*2048 会超限
     if model.startswith("wan2.6"):
-        return "1152*1536" if ratio == "3:4" else "1440*1440"
+        return {"3:4": "1152*1536", "4:5": "1280*1600"}.get(ratio, "1440*1440")
     if ratio == "3:4":
         return "1536*2048" if is_wan else "768*1024"
+    if ratio == "4:5":
+        return "1536*1920" if is_wan else "960*1200"
     return "2048*2048" if is_wan else "1024*1024"
 
 _DOUBAO_ANTI_AI = (
@@ -520,7 +522,8 @@ async def _pcore_request(s, prompt: str, ref_images: list[str], size: str, model
     kind = "edits" if ref_images else "generations"
     body: dict = {"model": model, "prompt": prompt, "n": 1, "size": size, "async": True}
     if ref_images:
-        # pcore 的 edits 要求 images 为 [{image_url: ...}] 对象数组（实测纯字符串报 images[].image_url is required）
+        # Pcore 的 gpt-image-2 通道实际要求 OpenAI 兼容对象数组。
+        # 虽然通用文档描述字符串数组也可用，该通道会在异步任务中校验 images[].image_url。
         body["images"] = [{"image_url": await _resolve_ref(u)} for u in ref_images]
 
     async with httpx.AsyncClient(timeout=max(s.request_timeout, 600), verify=False) as client:
@@ -723,6 +726,7 @@ async def advance_generation(
             kind = "edits" if ref_images else "generations"
             body: dict = {"model": model, "prompt": prompt, "n": 1, "size": size, "async": True}
             if ref_images:
+                # 与 _pcore_request 保持同一经 gpt-image-2 实际验证的形状，避免 Worker 路径漂移。
                 body["images"] = [{"image_url": await _resolve_ref(item)} for item in ref_images]
             if edit_mask:
                 # Pcore 的图片参数使用 JSON {image_url}；蒙版沿用相同的 OpenAI 兼容对象形式。

server/services/job_worker.py

@@ -6,6 +6,8 @@ import logging
 import socket
 import uuid
 
+import httpx
+
 from config import get_settings
 from db import get_session_factory
 from models import ProductAsset
@@ -23,6 +25,7 @@ from services.jobs import (
 )
 from services.prompts import build_context, build_prompt
 from services.prompts.common import requirements_block
+from services.prompts.single_edit import build_single_edit_prompt
 from services.storage import get_storage
 from services.suite_service import append_generated_asset
 from services.tasks import TaskImage
@@ -31,6 +34,10 @@ from services.video_renderer import VideoRenderCancelled, render_slideshow
 
 log = logging.getLogger("background.worker")
 
+# 异步供应商任务已经持久化了 task id 后，轮询网络失败不能把任务判为失败。
+# 下一次轮询复用同一 task id，不会再次提交/计费。
+PROVIDER_POLL_RETRY_DELAY_SECONDS = 30
+
 
 def _worker_owner() -> str:
     return f"{socket.gethostname()}-{uuid.uuid4().hex[:10]}"
@@ -191,7 +198,11 @@ async def process_item(claim: ClaimedItem) -> None:
         "pcore": "gpt-image-2",
     }.get(provider, settings.ark_image_model)
     try:
-        refs = _refs_for_job(list(payload.get("ref_images") or []), job_item)
+        single_edit = payload.get("single_edit")
+        refs = (
+            [image["url"] for image in payload.get("ref_images") or []]
+            if single_edit else _refs_for_job(list(payload.get("ref_images") or []), job_item)
+        )
         operation = payload.get("operation", "edit")
         target_size = _image_size(provider, claim.ratio, is_wan=provider == "tongyi" and model.lower().startswith("wan"), model=model)
         edit_mask: str | None = None
@@ -207,12 +218,21 @@ async def process_item(claim: ClaimedItem) -> None:
                 edit_mask = outpaint.mask_data_url
         else:
             context = build_context(payload.get("context") or {}, fallback_name="")
-            prompt_extra = job_item.get("prompt_extra") or job_item
-            prompt = build_prompt(
-                provider, model, job_item.get("kind", "custom"), context, claim.style_set, claim.lang,
-                extra=prompt_extra, style_prompt=payload.get("style_prompt"), requirements=payload.get("requirements"),
-                requirements_exclusive=bool(payload.get("requirements_exclusive")),
-            )
+            if single_edit:
+                prompt = build_single_edit_prompt(
+                    requirements=payload.get("requirements") or "", config=single_edit,
+                    context=context, style_set=claim.style_set,
+                    style_prompt=payload.get("style_prompt"), ratio=claim.ratio,
+                    is_size_chart=job_item.get("kind") == "size_chart",
+                    reference_count=len(refs) - 1,
+                )
+            else:
+                prompt_extra = job_item.get("prompt_extra") or job_item
+                prompt = build_prompt(
+                    provider, model, job_item.get("kind", "custom"), context, claim.style_set, claim.lang,
+                    extra=prompt_extra, style_prompt=payload.get("style_prompt"), requirements=payload.get("requirements"),
+                    requirements_exclusive=bool(payload.get("requirements_exclusive")),
+                )
         is_wan = provider == "tongyi" and model.lower().startswith("wan")
         result = await advance_generation(
             provider, prompt, refs,
@@ -274,6 +294,21 @@ async def process_item(claim: ClaimedItem) -> None:
             claim.item_id, result_url=result_url, source_url=source_url or None,
             watermark_config=watermark_config, generated_asset_id=asset_id,
         )
+    except (httpx.TimeoutException, httpx.TransportError) as exc:
+        if claim.provider_task_id and claim.provider_task_kind:
+            log.warning(
+                "供应商轮询网络暂时失败，%ds 后重试同一任务 %s：%s",
+                PROVIDER_POLL_RETRY_DELAY_SECONDS, claim.provider_task_id, exc.__class__.__name__,
+            )
+            await mark_waiting_provider(
+                claim.item_id,
+                claim.provider_task_id,
+                claim.provider_task_kind,
+                delay_seconds=PROVIDER_POLL_RETRY_DELAY_SECONDS,
+            )
+            return
+        log.exception("后台图片任务 %s/%s 网络请求失败", claim.job_id, claim.item_id)
+        await mark_item_failed(claim.item_id, str(exc) or exc.__class__.__name__)
     except Exception as exc:  # noqa: BLE001
         log.exception("后台图片任务 %s/%s 执行失败", claim.job_id, claim.item_id)
         await mark_item_failed(claim.item_id, str(exc))

server/services/prompts/common.py

@@ -140,7 +140,7 @@ def build_context(raw: dict, fallback_name: str = "", fallback_desc: str = "") -
             (label, dimensions[key]) for key, label in (("length", "length"), ("width", "width"), ("height", "height"))
             if dimensions.get(key) is not None
         ]
-    unit = (dimensions.get("unit") or "mm") if isinstance(dimensions, dict) else "mm"
+    unit = (dimensions.get("unit") or "cm") if isinstance(dimensions, dict) else "cm"
     dimensions_line = "; ".join(f"{label}: {value} {unit}" for label, value in known_dimensions)
     known_labels = [label for label, _ in known_dimensions]
     missing_labels = [label for label in ("length", "width", "height") if label not in known_labels]

server/services/prompts/single_edit.py

@@ -0,0 +1,68 @@
+"""单张 AI 编辑提示词：一次图像模型调用完成编辑、翻译和排版。"""
+from __future__ import annotations
+
+from services.prompts.common import requirements_block, resolve_style
+
+
+LANGUAGES = {"ru": "Russian", "ar": "Arabic", "en": "English"}
+
+BACKGROUNDS = {
+    "auto": "Keep the original background unless the user's requirements ask to change it.",
+    "white": "Replace only the background with clean pure white; retain natural contact shadows.",
+    "solid": "Replace only the background with a clean solid color as described below.",
+    "scene": "Replace only the background with the described realistic scene. Keep the product as the hero.",
+    "reference": "Use Image 2 for background and lighting direction. Do not copy its product.",
+}
+
+
+def build_single_edit_prompt(
+    *, requirements: str, config: dict, context: dict, style_set: int,
+    style_prompt: str | None, ratio: str, is_size_chart: bool, reference_count: int,
+) -> str:
+    """Image 1 always supplies the product; following images are supporting references only."""
+    language = LANGUAGES.get(config.get("target_language"), "Russian")
+    background = config.get("background") or "auto"
+    sections = [
+        requirements_block(requirements),
+        "Create exactly ONE finished e-commerce image. Image 1 is the fixed main product image. "
+        "Preserve the product's shape, proportions, color, material, logo and distinguishing details. "
+        "Do not replace it with products from any other image.",
+    ]
+    if reference_count:
+        sections.append(
+            f"Images 2 through {reference_count + 1} are auxiliary references only. "
+            "Use them for product detail, material, scene or lighting as appropriate; "
+            "Image 1 remains the source of truth for the product."
+        )
+    sections.append("BACKGROUND: " + BACKGROUNDS.get(background, BACKGROUNDS["auto"]))
+    if config.get("background_description"):
+        sections.append("Background details: " + str(config["background_description"]).strip())
+    if config.get("translate_text"):
+        sections.append(
+            f"In this same image-generation pass, read the visible marketing text in Image 1, "
+            f"translate its meaning into {language}, replace the source-language marketing copy, "
+            "and redesign the text layout for a polished, legible composition. "
+            "For Arabic, use correct right-to-left typesetting. Keep factual values and units exact. "
+            "Preserve text printed on the physical product and its branding unless the user explicitly asks to edit them. "
+            "Do not add unsupported claims."
+        )
+    else:
+        sections.append(
+            "Do not invent marketing copy or translate existing image text. "
+            "Preserve text printed on the physical product unless explicitly requested otherwise."
+        )
+    if is_size_chart:
+        annotation = context.get("dimension_annotation_instruction")
+        if annotation:
+            sections.append(
+                "Create a clean product measurement diagram with legible dimension arrows. "
+                + str(annotation)
+            )
+    style = resolve_style(style_set, style_prompt)
+    sections.append("VISUAL STYLE: " + str(style["tone"]))
+    sections.append(
+        f"OUTPUT: exactly {ratio} portrait or square composition as specified. "
+        "Apply the user's explicit background and text instructions before style preferences. "
+        "Do not add a watermark or extra panels."
+    )
+    return "\n\n".join(part for part in sections if part)