app.py

import gradio as gr
import torch
from diffusers import StableDiffusionImg2ImgPipeline
from PIL import Image

# --- Model Loading --- #
print('Loading Ghibli and Anime style models...')

# Determine the device to run the model on
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f"Using device: {device}")

# --- Load Ghibli Style Model ---
ghibli_model_id = 'nitrosocke/Ghibli-Diffusion'

# Note: `torch_dtype` is used here instead of `dtype` due to observed behavior in diffusers 0.36.0
# where `dtype` was ignored and `torch_dtype` was effective despite deprecation warnings.
if device == 'cuda':
    ghibli_pipeline = StableDiffusionImg2ImgPipeline.from_pretrained(
        ghibli_model_id,
        torch_dtype=torch.float16,
        cache_dir='./model_cache_ghibli'
    )
else:
    ghibli_pipeline = StableDiffusionImg2ImgPipeline.from_pretrained(
        ghibli_model_id,
        torch_dtype=torch.float32,
        cache_dir='./model_cache_ghibli'
    )
ghibli_pipeline.to(device)
print(f'Ghibli Style Model ({ghibli_model_id}) loaded successfully.')

# --- Load Anime Style Model ---
anime_model_id = 'hakurei/waifu-diffusion'
if device == 'cuda':
    anime_pipeline = StableDiffusionImg2ImgPipeline.from_pretrained(
        anime_model_id,
        torch_dtype=torch.float16,
        cache_dir='./model_cache_anime'
    )
else:
    anime_pipeline = StableDiffusionImg2ImgPipeline.from_pretrained(
        anime_model_id,
        torch_dtype=torch.float32,
        cache_dir='./model_cache_anime'
    )
anime_pipeline.to(device)
print(f'Anime Style Model ({anime_model_id}) loaded successfully.')

print('Both Ghibli and Anime Style Models loaded and moved to device successfully.')

# --- Transformation Function ---
def cartoon_transform(input_image: Image.Image, style: str) -> Image.Image:
    """
    Applies a cartoon-style transformation to the input image using the loaded Stable Diffusion pipelines.

    Args:
        input_image (PIL.Image.Image): The input image to transform.
        style (str): The desired cartoon style ('Ghibli' or 'Anime').

    Returns:
        PIL.Image.Image: The transformed image in the selected cartoon style.
    """
    # Ensure the image is in RGB format
    if input_image.mode != 'RGB':
        input_image = input_image.convert('RGB')

    # Set reasonable dimensions to avoid excessive memory usage and ensure reasonable processing time
    # while maintaining aspect ratio
    max_dim = 768 # Maximum dimension for processing
    width, height = input_image.size
    if max(width, height) > max_dim:
        ratio = max_dim / max(width, height)
        new_width = int(width * ratio)
        new_height = int(height * ratio)
        input_image = input_image.resize((new_width, new_height), Image.LANCZOS)

    # Define pipelines, prompts, and parameters based on style
    if style == 'Ghibli':
        pipeline_to_use = ghibli_pipeline
        prompt = "Studio Ghibli style, detailed, vibrant colors, fantasy, magical, serene"
        strength = 0.75
        guidance_scale = 7.5
        num_inference_steps = 25 # Reduced for faster processing
    elif style == 'Anime':
        pipeline_to_use = anime_pipeline
        prompt = "anime character, vibrant, digital art, high quality, detailed eyes"
        strength = 0.8 # Slightly higher strength for a more pronounced anime effect
        guidance_scale = 8.0
        num_inference_steps = 25 # Reduced for faster processing
    else:
        raise ValueError(f"Unsupported style: {style}. Choose from 'Ghibli' or 'Anime'.")

    # Run the image-to-image pipeline
    transformed_image = pipeline_to_use(
        prompt=prompt,
        image=input_image,
        strength=strength,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps
    ).images[0]

    print(f'Image transformed to {style} style.')
    return transformed_image

# --- Gradio Interface ---
# Available cartoon styles
cartoon_styles = ['Ghibli', 'Anime']

# Create the Gradio interface
iface = gr.Interface(
    fn=cartoon_transform,
    inputs=[
        gr.Image(type='pil', label='Upload your picture'),
        gr.Dropdown(
            choices=cartoon_styles,
            label='Select Cartoon Style',
            value='Ghibli' # Default selection
        )
    ],
    outputs=gr.Image(type='pil', label='Transformed Image'),
    title='Cartoon Style Image Transformer',
    description='Upload a picture and transform it into various cartoon styles.'
)

# Launch the Gradio app - this part is typically removed or commented out when deploying to Hugging Face Spaces,
# as Spaces handle the launch automatically.
if __name__ == '__main__':
    print('Launching Gradio interface locally...')
    iface.launch(share=True) # share=True for Colab, change to False for local dev