From caec23b5cc1f8975e1224fc6d23b7760f53f53d8 Mon Sep 17 00:00:00 2001
From: Jony <951565127@qq.com>
Date: Sun, 26 Apr 2026 01:13:13 +0800
Subject: [PATCH] =?UTF-8?q?feat:=20GPT=20Image=20=E8=AF=B7=E6=B1=82?=
 =?UTF-8?q?=E4=BD=93=E8=B6=85=2020MB=20=E6=97=B6=E8=87=AA=E5=8A=A8?=
 =?UTF-8?q?=E7=AD=89=E6=AF=94=E7=BC=A9=E6=94=BE=E5=9B=BE=E5=83=8F?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 clients/gpt_image_client.py | 59 +++++++++++++++++++++++++++++++++++--
 1 file changed, 57 insertions(+), 2 deletions(-)

diff --git a/clients/gpt_image_client.py b/clients/gpt_image_client.py
index 23fa308..de9012f 100644
--- a/clients/gpt_image_client.py
+++ b/clients/gpt_image_client.py
@@ -70,6 +70,41 @@ class GptImageClient:
 
     # ── 图像转换工具 ──────────────────────────────────────────────────────────
 
+    # ── 请求体大小限制 ────────────────────────────────────────────────────────
+    _MAX_BODY_BYTES = 20 * 1024 * 1024   # 20 MB
+
+    @staticmethod
+    def _shrink_png_to_limit(png_bytes: bytes, max_bytes: int, label: str = "") -> bytes:
+        """
+        若 PNG bytes 超过 max_bytes，按等比缩放反复压缩直到满足限制。
+        每次将面积缩小至约 80%（线性尺寸缩小至约 89.4%）。
+        """
+        if len(png_bytes) <= max_bytes:
+            return png_bytes
+
+        img = Image.open(BytesIO(png_bytes))
+        w, h = img.size
+        original_size = len(png_bytes)
+        step = 0
+
+        while len(png_bytes) > max_bytes:
+            scale = 0.894   # sqrt(0.8)，面积缩小 20%
+            w = max(1, int(w * scale))
+            h = max(1, int(h * scale))
+            img = img.resize((w, h), Image.LANCZOS)
+            buf = BytesIO()
+            img.save(buf, format="PNG")
+            png_bytes = buf.getvalue()
+            step += 1
+
+        tag = f" ({label})" if label else ""
+        print(
+            f"[o1key GPT Image] 图像{tag}超出 {max_bytes // (1024*1024)}MB 限制，"
+            f"已等比缩放 {step} 次：{original_size // 1024}KB → {len(png_bytes) // 1024}KB "
+            f"（{w}×{h}）"
+        )
+        return png_bytes
+
     @staticmethod
     def _tensor_to_png_bytes(tensor: torch.Tensor) -> bytes:
         """
@@ -220,8 +255,20 @@ class GptImageClient:
         if image_tensor is not None:
             pil_images = tensor_to_pil(image_tensor)
             data_urls = []
-            for img in pil_images:
-                b64 = encode_image_to_base64(img, format="PNG")
+            for idx_img, img in enumerate(pil_images):
+                buf = BytesIO()
+                img.save(buf, format="PNG")
+                png_bytes = buf.getvalue()
+                # 单张图像预算：20MB 按图数平摊，至少保留 1MB 给其他字段
+                per_image_budget = max(
+                    1024 * 1024,
+                    (self._MAX_BODY_BYTES - 1024 * 1024) // len(pil_images),
+                )
+                # base64 膨胀约 4/3，所以 PNG 目标上限 = budget * 3/4
+                png_budget = int(per_image_budget * 3 / 4)
+                label = f"第{idx_img + 1}张" if len(pil_images) > 1 else ""
+                png_bytes = self._shrink_png_to_limit(png_bytes, png_budget, label)
+                b64 = base64.b64encode(png_bytes).decode("utf-8")
                 data_urls.append(f"data:image/png;base64,{b64}")
             body["image"] = data_urls[0] if len(data_urls) == 1 else data_urls
             mode = f"图生图（参考图 {len(data_urls)} 张）"
@@ -302,9 +349,17 @@ class GptImageClient:
         form.add_field("size", size if size else "auto")
 
         # 多图：用 image[] 数组字段逐张附加，支持 gpt-image-1.5 最多 16 张
+        # 预算：20MB 按图数平摊，蒙版预留 1MB
+        mask_reserve = 1024 * 1024 if mask_tensor is not None else 0
+        per_image_budget = max(
+            1024 * 1024,
+            (self._MAX_BODY_BYTES - mask_reserve) // num_images,
+        )
         for i in range(num_images):
             frame = image_tensor[i:i+1]               # [1,H,W,C]
             img_bytes = self._tensor_to_png_bytes(frame)
+            label = f"第{i + 1}张" if num_images > 1 else ""
+            img_bytes = self._shrink_png_to_limit(img_bytes, per_image_budget, label)
             form.add_field(
                 "image[]",
                 img_bytes,