Upload folder using huggingface_hub

gradio_sketch2image.py

@@ -64,29 +64,54 @@ def pil_image_to_data_uri(img, format="PNG"):
     return f"data:image/{format.lower()};base64,{img_str}"
 
 
+def normalize_image(image, range_from=(-1, 1)):
+    """
+    Normalize the input image to a specified range.
+    
+    :param image: The PIL Image to be normalized.
+    :param range_from: The target range for normalization, typically (-1, 1) or (0, 1).
+    :return: Normalized image tensor.
+    """
+    # Convert the image to a tensor
+    image_t = F.to_tensor(image)
+    
+    if range_from == (-1, 1):
+        # Normalize from [0, 1] to [-1, 1]
+        image_t = image_t * 2 - 1
+    
+    return image_t
+
+
 def run(image, prompt, prompt_template, style_name, seed, val_r):
     print(f"prompt: {prompt}")
     print("sketch updated")
+    print(image)
     if image is None:
-        ones = Image.new("L", (512, 512), 255)
+        ones = Image.new("L", (1024, 1024), 255)
         temp_uri = pil_image_to_data_uri(ones)
         return ones, gr.update(link=temp_uri), gr.update(link=temp_uri)
+    print(f"Input Image Size: {image.size}")
     prompt = prompt_template.replace("{prompt}", prompt)
     image = image.convert("RGB")
-    image_t = F.to_tensor(image) > 0.5
-    image_t = F.to_tensor(image) * 2 - 1  # Normalize to [-1, 1]
+   
+    # Normalize image to [-1, 1]
+    image_t = normalize_image(image, range_from=(-1, 1))
+    
     print(f"r_val={val_r}, seed={seed}")
     
     with torch.no_grad():
         c_t = image_t.unsqueeze(0).cuda().float()
         torch.manual_seed(seed)
         B, C, H, W = c_t.shape
-        noise = torch.randn((1, 4, H // 8, W // 8), device=c_t.device)
+        #noise = torch.randn((1, 4, H // 8, W // 8), device=c_t.device)
+        noise = torch.randn((1, 4, H // 8, W // 8), device=c_t.device) * 0.8 
         print("Calling Pix2Pix model... ct: {}, prompt: {}, deterministic: False, r: {}, noise_map: {}".format(c_t.shape, prompt, val_r, noise.shape))  
         output_image = model(c_t, prompt, deterministic=False, r=val_r, noise_map=noise)
     output_pil = F.to_pil_image(output_image[0].cpu() * 0.5 + 0.5)
     input_sketch_uri = pil_image_to_data_uri(Image.fromarray(255 - np.array(image)))
     output_image_uri = pil_image_to_data_uri(output_pil)
+    
+    print(f"Output Image Size: {output_pil.size}")
     return (
         output_pil,
         gr.update(link=input_sketch_uri),

gradio_sketch2imagehd.py

@@ -95,6 +95,62 @@ def extract_main_words(item: str) -> str:
     nouns = [word.capitalize() for word, tag in tagged if tag in ('NN', 'NNP', 'NNPS', 'NNS')]
     return ' '.join(nouns)
 
+def normalize_image(image, range_from=(-1, 1)):
+    """
+    Normalize the input image to a specified range.
+    
+    :param image: The PIL Image to be normalized.
+    :param range_from: The target range for normalization, typically (-1, 1) or (0, 1).
+    :return: Normalized image tensor.
+    """
+    # Convert the image to a tensor
+    image_t = F.to_tensor(image)
+    
+    if range_from == (-1, 1):
+        # Normalize from [0, 1] to [-1, 1]
+        image_t = image_t * 2 - 1
+    
+    return image_t
+
+
+def run(image, prompt, prompt_template, style_name, seed, val_r):
+    print(f"prompt: {prompt}")
+    print("sketch updated")
+    print(image)
+    if image is None:
+        ones = Image.new("L", (1024, 1024), 255)
+        temp_uri = pil_image_to_data_uri(ones)
+        return ones, gr.update(link=temp_uri), gr.update(link=temp_uri)
+    print(f"Input Image Size: {image.size}")
+    prompt = prompt_template.replace("{prompt}", prompt)
+    image = image.convert("RGB")
+    image_t = normalize_image(image, range_from=(-1, 1))
+    
+    print(f"r_val={val_r}, seed={seed}")
+    
+    with torch.no_grad():
+        c_t = image_t.unsqueeze(0).cuda().float()
+        torch.manual_seed(seed)
+        #B, C, H, W = c_t.shape
+        #noise = torch.randn((1, 4, H // 8, W // 8), device=c_t.device)
+        #noise = torch.randn((1, 4, H // 8, W // 8), device=c_t.device) * 0.8 
+
+        B, C, H, W = c_t.shape
+        #noise = torch.randn((1, 4, H // 8, W // 8), device=c_t.device)
+        noise = torch.randn((1, 4, H // 8, W // 8), device=c_t.device) * 0.4 
+        print("Calling Pix2Pix model... ct: {}, prompt: {}, deterministic: False, r: {}, noise_map: {}".format(c_t.shape, prompt, val_r, noise.shape))  
+        output_image = pix2pix_model(c_t, prompt, deterministic=False, r=val_r, noise_map=noise)
+    output_pil = F.to_pil_image(output_image[0].cpu() * 0.5 + 0.5)
+    input_sketch_uri = pil_image_to_data_uri(Image.fromarray(255 - np.array(image)))
+    output_image_uri = pil_image_to_data_uri(output_pil)
+    
+    print(f"Output Image Size: {output_pil.size}")
+    return (
+        output_pil,
+        gr.update(link=input_sketch_uri),
+        gr.update(link=output_image_uri),
+    )
+
 def run(image, prompt, prompt_template, style_name, seed, val_r):
     """Runs the main image processing pipeline."""
     logging.debug("Running model inference...")

run - Copy (2).cmd

@@ -0,0 +1,12 @@
+@echo off
+REM Activate the Conda environment
+call conda activate img2img-turbo
+
+REM Navigate to the project directory
+cd /d D:\I+D\ia\img2img\img2img-turbo
+
+REM Run the Python script
+python gradio_sketch2image.py
+
+REM Pause the script to keep the command prompt open in case of errors
+pause