Update app.py

app.py

@@ -80,7 +80,6 @@ class OrangeRedTheme(Soft):
 orange_red_theme = OrangeRedTheme()
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-
 print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES"))
 print("torch.__version__ =", torch.__version__)
 print("Using device:", device)
@@ -110,29 +109,29 @@ except Exception as e:
 
 # 不加 VAE 显存优化。因为 VAE Tiling/Slicing 会增加生成耗时，进而消耗你的 ZeroGPU Quota。
 # A100 处理 1024 分辨率足够，所以全速运行即可。
-
 MAX_SEED = np.iinfo(np.int32).max
 
-def update_dimensions_on_upload(image, max_size=768):
+def update_dimensions_on_upload(image, max_size=768, multiple=32):
+    """
+    修改点：将原先的向下取整为 8 的倍数，改为四舍五入到 32 的倍数。
+    这彻底阻止了 Diffusers 模型由于不匹配 DiT patch 结构而产生的暗中裁剪。
+    """
     if image is None:
         return max_size, max_size
-
     original_width, original_height = image.size
 
     # Optimization: Do not upscale small images unnecessarily
     if original_width <= max_size and original_height <= max_size:
         new_width, new_height = original_width, original_height
-    elif original_width > original_height:
-        new_width = max_size
-        aspect_ratio = original_height / original_width
-        new_height = int(new_width * aspect_ratio)
     else:
-        new_height = max_size
-        aspect_ratio = original_width / original_height
-        new_width = int(new_height * aspect_ratio)
+        # Scale down proportionally based on the longest edge
+        scale = min(max_size / original_width, max_size / original_height)
+        new_width = int(original_width * scale)
+        new_height = int(original_height * scale)
 
-    new_width = (new_width // 8) * 8
-    new_height = (new_height // 8) * 8
+    # 强制将宽高锁定为 `multiple` (即 32) 的倍数
+    new_width = max(multiple, round(new_width / multiple) * multiple)
+    new_height = max(multiple, round(new_height / multiple) * multiple)
 
     return new_width, new_height
 
@@ -189,12 +188,13 @@ def process_and_concat(images, prompt, seed, randomize_seed, guidance_scale, ste
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
 
-    # --- NEW: Explicitly resize to avoid Diffusers' internal CenterCrop ---
+    # --- 避免裁剪的关键 ---
+    # CPU 端直接把原图 resize 到严格兼容模型底层的尺寸 (32的倍数)。
+    # 不依赖 Diffusers 管道内部做任何尺寸处理。
     pre_resized_image = orig_image.resize((width, height), Image.Resampling.LANCZOS)
 
     # --- Step 2: Dispatch to GPU ---
     # GPU block is completely clean, measuring only pure generation time
-    # Pass 'pre_resized_image' instead of 'orig_image'
     result_image = run_gpu_inference(
         pre_resized_image, prompt, seed, guidance_scale, steps, width, height
     )
@@ -242,7 +242,7 @@ css = """
 
 with gr.Blocks() as demo:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown("# **FireRed-Image-Edit-1.0-Fast**", elem_id="main-title")
+        gr.Markdown("# FireRed-Image-Edit-1.0-Fast", elem_id="main-title")
         gr.Markdown("Perform image edits using [FireRed-Image-Edit-1.0](https://huggingface.co/FireRedTeam/FireRed-Image-Edit-1.0) with 4-step fast inference. Open on [GitHub](https://github.com/PRITHIVSAKTHIUR/FireRed-Image-Edit-1.0-Fast)")
 
         with gr.Row(equal_height=True):
@@ -255,20 +255,17 @@ with gr.Blocks() as demo:
                     height=300,
                     allow_preview=True
                 )
-
                 prompt = gr.Text(
                     label="Edit Prompt",
                     show_label=True,
                     max_lines=2,
                     placeholder="e.g., transform into anime, upscale, change lighting...",
                 )
-
                 run_button = gr.Button("Edit Image", variant="primary")
 
             with gr.Column():
                 output_image = gr.Image(interactive=False, format="png")
                 output_file = gr.File(label="Download Lossless Merged Image")
-
                 with gr.Accordion("Advanced Settings", open=False, visible=False):
                     seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
                     randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)