app.py

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

# Load pre-trained models for CPU
print("Loading models...")

# Stable Diffusion for colorization and inpainting
colorization_pipeline = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float32)
colorization_pipeline = colorization_pipeline.to("cpu")

# Img2Img pipeline for denoising
denoising_pipeline = StableDiffusionImg2ImgPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", torch_dtype=torch.float32)
denoising_pipeline = denoising_pipeline.to("cpu")

# Upscaling pipeline using stabilityai's upscaler
upscaling_pipeline = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", torch_dtype=torch.float32)
upscaling_pipeline = upscaling_pipeline.to("cpu")

# Define image processing functions
def colorize_image(grayscale_image):
    grayscale_image = grayscale_image.convert("RGB").resize((512, 512))
    prompt = "A colorful and realistic version of this grayscale photo."
    result = colorization_pipeline(prompt, image=grayscale_image)
    return result.images[0]

def denoise_image(noisy_image, strength: float = 0.7):
    noisy_image = noisy_image.convert("RGB").resize((512, 512))
    prompt = "A high-quality version of this image."
    result = denoising_pipeline(prompt, image=noisy_image, strength=strength)
    return result.images[0]

def inpaint_image(base_image, mask_image):
    if mask_image is None:
        raise ValueError("Mask image is required for inpainting.")
    base_image = base_image.convert("RGB").resize((512, 512))
    mask_image = mask_image.convert("RGB").resize((512, 512))
    prompt = "A completed and visually realistic image with the missing parts filled in."
    result = colorization_pipeline(prompt, image=base_image, mask_image=mask_image)
    return result.images[0]

def upscale_image(low_res_image):
    low_res_image = low_res_image.convert("RGB").resize((256, 256))  # Upscaler expects smaller inputs
    prompt = "An upscaled version of this image."
    result = upscaling_pipeline(prompt, image=low_res_image)
    return result.images[0]

# Main function to process based on the selected task
def process_image(task, image, mask=None):
    if task == "Colorize":
        return colorize_image(image)
    elif task == "Denoise":
        return denoise_image(image)
    elif task == "Inpaint":
        return inpaint_image(image, mask)
    elif task == "Upscale":
        return upscale_image(image)
    else:
        raise ValueError("Invalid task selected.")

# Gradio Interface
interface = gr.Interface(
    fn=process_image,
    inputs=[
        gr.Dropdown(
            ["Colorize", "Denoise", "Inpaint", "Upscale"],
            label="Select Task",
            value="Colorize"
        ),
        gr.Image(type="pil", label="Upload Image"),
        gr.Image(type="pil", label="Upload Mask (for Inpainting)", value=None)  # Allow empty input for mask
    ],
    outputs=gr.Image(label="Processed Image"),
    title="AI-Powered Image Restoration",
    description=(
        "Select an image restoration task from the dropdown menu:\n\n"
        "- **Colorize**: Convert grayscale images to realistic color.\n"
        "- **Denoise**: Enhance image quality by removing noise.\n"
        "- **Inpaint**: Fill missing parts of an image using a mask.\n"
        "- **Upscale**: Increase image resolution while preserving details."
    )
)

if __name__ == "__main__":
    interface.launch(share=True)