feat: 试用过程中的细节优化

server/services/generator.py

@@ -10,6 +10,7 @@ from __future__ import annotations
 
 import asyncio
 import base64
+import io
 import logging
 import mimetypes
 import re
@@ -17,6 +18,7 @@ from dataclasses import dataclass
 from typing import Awaitable, Callable
 
 import httpx
+from PIL import Image, ImageOps
 
 from config import get_settings
 from services.prompts import build_context, build_prompt, type_name
@@ -84,8 +86,8 @@ def _image_size(provider: str, ratio: str, is_wan: bool = True, model: str = "")
         # 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"
     if provider == "pcore":
-        # pcore 的 size 是枚举（默认 1024x1024），不接受自定义尺寸；用 gpt-image 枚举：竖版 1024x1536（近似 3:4）
-        return "1024x1536" if ratio == "3:4" else "1024x1024"
+        # 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"
     # 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"):
@@ -141,6 +143,75 @@ async def _resolve_ref(url: str) -> str:
     return _bytes_to_data_uri(data, mime)
 
 
+@dataclass(frozen=True)
+class OutpaintCanvas:
+    """用于受限扩图的预处理结果。
+
+    ``image_data_url`` 与 ``mask_data_url`` 尺寸严格一致：蒙版不透明的区域是原图，
+    透明区域才允许供应商生成。最终仍会把 ``original`` 覆盖回结果，防止供应商在
+    蒙版区域有轻微重绘。
+    """
+
+    image_data_url: str
+    mask_data_url: str
+    original: Image.Image
+    box: tuple[int, int, int, int]
+    size: tuple[int, int]
+
+    def restore_original(self, generated: bytes) -> bytes:
+        """恢复原图区域，确保输出中原图像素不受模型影响。"""
+        with Image.open(io.BytesIO(generated)) as result:
+            result = ImageOps.exif_transpose(result).convert("RGBA")
+            if result.size != self.size:
+                # 供应商偶发不遵从 size 时仍输出固定 3:4；随后会精确覆盖原图区域。
+                result = ImageOps.fit(result, self.size, method=Image.Resampling.LANCZOS)
+            result.alpha_composite(self.original, self.box[:2])
+            out = io.BytesIO()
+            result.save(out, format="PNG")
+            return out.getvalue()
+
+
+def _png_data_url(image: Image.Image) -> str:
+    out = io.BytesIO()
+    image.save(out, format="PNG")
+    return _bytes_to_data_uri(out.getvalue(), "image/png")
+
+
+async def prepare_outpaint_canvas(url: str, size: str) -> OutpaintCanvas:
+    """把单张参考图置于目标画布，并生成仅开放新增边缘的 PNG alpha 蒙版。"""
+    try:
+        target_width, target_height = (int(value) for value in size.lower().split("x", 1))
+    except (TypeError, ValueError) as exc:
+        raise ValueError(f"仅扩图需要合法的宽高尺寸，收到: {size}") from exc
+    if target_width <= 0 or target_height <= 0:
+        raise ValueError(f"仅扩图需要正数宽高尺寸，收到: {size}")
+
+    data, _ = await _resolve_ref_bytes(url)
+    with Image.open(io.BytesIO(data)) as opened:
+        source = ImageOps.exif_transpose(opened).convert("RGBA")
+        if source.width <= 0 or source.height <= 0:
+            raise ValueError("参考底图尺寸无效")
+        scale = min(target_width / source.width, target_height / source.height)
+        source_width = max(1, round(source.width * scale))
+        source_height = max(1, round(source.height * scale))
+        original = source.resize((source_width, source_height), Image.Resampling.LANCZOS)
+
+    x = (target_width - original.width) // 2
+    y = (target_height - original.height) // 2
+    canvas = Image.new("RGBA", (target_width, target_height), (0, 0, 0, 0))
+    canvas.alpha_composite(original, (x, y))
+    # OpenAI/GPT Image 兼容蒙版约定：透明像素是可编辑区域；原图区域不透明。
+    mask = Image.new("RGBA", (target_width, target_height), (0, 0, 0, 0))
+    mask.paste((255, 255, 255, 255), (x, y, x + original.width, y + original.height))
+    return OutpaintCanvas(
+        image_data_url=_png_data_url(canvas),
+        mask_data_url=_png_data_url(mask),
+        original=original,
+        box=(x, y, x + original.width, y + original.height),
+        size=(target_width, target_height),
+    )
+
+
 # ── Provider：豆包 Seedream（火山方舟）────────────────────────────────────
 
 async def generate_doubao(prompt: str, ref_images: list[str], size: str = "2048x2048", model: str | None = None) -> bytes:
@@ -562,6 +633,7 @@ async def advance_generation(
     model: str,
     provider_task_id: str | None = None,
     provider_task_kind: str | None = None,
+    edit_mask: str | None = None,
 ) -> GenerationAdvance:
     """推进一次图片生成，不在本函数内无限轮询。
 
@@ -569,6 +641,8 @@ async def advance_generation(
     能继续外部任务，且在等待期间可以领取其他任务。豆包和同步千问仍在一次调用内完成。
     """
     settings = get_settings()
+    if edit_mask and provider != "pcore":
+        raise RuntimeError("带蒙版的仅扩图当前仅支持 Pcore gpt-image-2")
     if provider == "doubao":
         return GenerationAdvance(data=await generate_doubao(prompt, ref_images, size=size, model=model))
 
@@ -650,6 +724,10 @@ async def advance_generation(
             body: dict = {"model": model, "prompt": prompt, "n": 1, "size": size, "async": True}
             if ref_images:
                 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。
+                # 透明区域允许编辑，不透明区域保留。若供应商轻微偏移，Worker 会二次覆盖原图。
+                body["mask"] = {"image_url": edit_mask}
             response = await client.post(
                 f"{base}/v1/images/{kind}", headers={**headers, "Content-Type": "application/json"}, json=body,
             )