fix: pad-to-bucket instead of stretch, preserve exact aspect ratio, edge-fill padding

app.py

@@ -380,23 +380,61 @@ def prepare_images_before_pipe(
     pil_images: List[Image.Image],
     allow_upscale: bool = UPSCALE_SMALL_IMAGES,
     divisible_by: int = 16,
-) -> Tuple[List[Image.Image], int, int]:
+) -> Tuple[List[Image.Image], int, int, tuple]:
+    """准备图片：等比缩放 + 补边到最佳 bucket，保留原始比例。
+    返回 (processed_images, width, height, pad_info)
+    pad_info = (pad_left, pad_top, content_w, content_h) 用于推理后裁剪补边。
+    """
     if not pil_images:
         raise ValueError("No input images.")
 
     base_w, base_h = pil_images[0].size
 
-    # 用 bucket 系统选择最佳推理分辨率（~1MP），匹配模型训练分辨率
-    # 避免原图分辨率过大/过小导致内容偏移或裁剪
-    target_w, target_h = pick_best_bucket(base_w, base_h, SAFE_BUCKETS, allow_upscale)
+    # 选最佳 bucket（~1MP，比例最接近）
+    bucket_w, bucket_h = pick_best_bucket(base_w, base_h, SAFE_BUCKETS, allow_upscale)
+
+    # 等比缩放 fit 到 bucket 内（不拉伸）
+    scale = min(bucket_w / base_w, bucket_h / base_h)
+    content_w = max(divisible_by, round(base_w * scale))
+    content_h = max(divisible_by, round(base_h * scale))
+
+    # 居中补边到 bucket 尺寸
+    pad_left = (bucket_w - content_w) // 2
+    pad_top = (bucket_h - content_h) // 2
+    pad_info = (pad_left, pad_top, content_w, content_h)
 
     processed = []
     for img in pil_images:
-        if img.size != (target_w, target_h):
-            img = img.resize((target_w, target_h), Image.LANCZOS)
-        processed.append(img)
-
-    return processed, target_w, target_h
+        # 等比缩放
+        resized = img.resize((content_w, content_h), Image.LANCZOS)
+        # 创建 bucket 大小的画布，边缘用镜像填充减少接缝
+        canvas = Image.new("RGB", (bucket_w, bucket_h), (0, 0, 0))
+        canvas.paste(resized, (pad_left, pad_top))
+
+        # 用边缘像素填充补边区域（比纯黑效果好）
+        import numpy as _np
+        arr = np.array(canvas)
+        res_arr = np.array(resized)
+        # 填充左右
+        if pad_left > 0:
+            left_col = res_arr[:, 0:1, :]
+            arr[pad_top:pad_top+content_h, :pad_left, :] = np.broadcast_to(left_col, (content_h, pad_left, 3))
+        right_start = pad_left + content_w
+        if right_start < bucket_w:
+            right_col = res_arr[:, -1:, :]
+            arr[pad_top:pad_top+content_h, right_start:, :] = np.broadcast_to(right_col, (content_h, bucket_w - right_start, 3))
+        # 填充上下
+        if pad_top > 0:
+            top_row = arr[pad_top:pad_top+1, :, :]
+            arr[:pad_top, :, :] = np.broadcast_to(top_row, (pad_top, bucket_w, 3))
+        bottom_start = pad_top + content_h
+        if bottom_start < bucket_h:
+            bottom_row = arr[bottom_start-1:bottom_start, :, :]
+            arr[bottom_start:, :, :] = np.broadcast_to(bottom_row, (bucket_h - bottom_start, bucket_w, 3))
+
+        processed.append(Image.fromarray(arr))
+
+    return processed, bucket_w, bucket_h, pad_info
 
 
 def extract_pil_from_source(source) -> Image.Image:
@@ -497,7 +535,7 @@ def infer(
 
     generator = torch.Generator(device=device).manual_seed(int(seed))
 
-    processed_images, width, height = prepare_images_before_pipe(
+    processed_images, width, height, pad_info = prepare_images_before_pipe(
         pil_images, allow_upscale=UPSCALE_SMALL_IMAGES
     )
 
@@ -519,7 +557,12 @@ def infer(
             true_cfg_scale=guidance_scale,
         ).images[0]
 
-        # ── 还原到原始尺寸（消除 16 对齐造成的裁剪）──
+        # ── 裁掉补边，还原到原始比例内容区域 ──
+        pad_left, pad_top, content_w, content_h = pad_info
+        if pad_left > 0 or pad_top > 0 or content_w < width or content_h < height:
+            result = result.crop((pad_left, pad_top, pad_left + content_w, pad_top + content_h))
+
+        # ── 还原到原始尺寸 ──
         if result.size != orig_size:
             result = result.resize(orig_size, Image.LANCZOS)