Update app.py

app.py

@@ -552,6 +552,9 @@ def process_input(input_image, upscale_factor, **kwargs):
 
     return input_image.resize((w, h)), w_original, h_original, was_resized
     
+from PIL import Image
+import numpy as np
+
 @spaces.GPU
 def infer_upscale(
     seed,
@@ -573,7 +576,7 @@ def infer_upscale(
 
         # rescale with upscale factor
         w, h = input_image.size
-        control_image = input_image.resize((w * upscale_factor, h * upscale_factor))
+        control_image = input_image.resize((w * upscale_factor, h * upscale_factor), Image.LANCZOS)
 
         generator = torch.Generator(device=device).manual_seed(seed)
 
@@ -588,24 +591,29 @@ def infer_upscale(
         # Convert to tensor and add batch dimension
         control_image = torch.from_numpy(np.array(control_image)).permute(2, 0, 1).float().unsqueeze(0).to(device) / 255.0
         
-        image = pipe_upscale(
-            prompt="",
-            control_image=control_image,
-            controlnet_conditioning_scale=controlnet_conditioning_scale,
-            num_inference_steps=num_inference_steps,
-            guidance_scale=3.5,
-            generator=generator,
-        ).images[0]
+        with torch.no_grad():
+            image = pipe_upscale(
+                prompt="",
+                control_image=control_image,
+                controlnet_conditioning_scale=controlnet_conditioning_scale,
+                num_inference_steps=num_inference_steps,
+                guidance_scale=3.5,
+                generator=generator,
+            ).images[0]
+
+        # Convert the image back to PIL Image
+        if isinstance(image, torch.Tensor):
+            image = image.cpu().permute(1, 2, 0).numpy()
+        
+        # Ensure the image data is in the correct range
+        image = np.clip(image * 255, 0, 255).astype(np.uint8)
+        image = Image.fromarray(image)
 
         if was_resized:
             gr.Info(
                 f"Resizing output image to targeted {w_original * upscale_factor}x{h_original * upscale_factor} size."
             )
-
-        # resize to target desired size
-        image = image.resize((w_original * upscale_factor, h_original * upscale_factor))
-        
-        # image is already a PIL Image, no need for further conversion
+            image = image.resize((w_original * upscale_factor, h_original * upscale_factor), Image.LANCZOS)
 
         return image, seed, num_inference_steps, upscale_factor, controlnet_conditioning_scale, gr.update(), gr.update(visible=False)
     except Exception as e:
@@ -715,10 +723,10 @@ with gr.Blocks(theme="Nymbo/Nymbo_Theme", css=css, delete_cache=(60, 3600)) as a
             controlnet_conditioning_scale = gr.Slider(
                 label="Controlnet Conditioning Scale",
                 minimum=0.1,
-                maximum=1.5,
-                step=0.1,
-                value=0.6,
-            )
+                maximum=1.0,
+                step=0.05,
+                value=0.5,  # 기본값을 0.5로 낮춤
+            )            
             upscale_seed = gr.Slider(
                 label="Seed for Upscaling",
                 minimum=0,