app.py

import torch
from PIL import Image
import numpy as np
from transformers import AutoImageProcessor, Swin2SRForImageSuperResolution
import gradio as gr
import spaces

def resize_image(image, max_size=2048):
    width, height = image.size
    if width > max_size or height > max_size:
        aspect_ratio = width / height
        if width > height:
            new_width = max_size
            new_height = int(new_width / aspect_ratio)
        else:
            new_height = max_size
            new_width = int(new_height * aspect_ratio)
        image = image.resize((new_width, new_height), Image.LANCZOS)
    return image

def split_image(image, chunk_size=512):
    width, height = image.size
    chunks = []
    for y in range(0, height, chunk_size):
        for x in range(0, width, chunk_size):
            chunk = image.crop((x, y, min(x + chunk_size, width), min(y + chunk_size, height)))
            chunks.append((chunk, x, y))
    return chunks

def stitch_image(chunks, original_size):
    result = Image.new('RGB', original_size)
    for img, x, y in chunks:
        result.paste(img, (x, y))
    return result

def upscale_chunk(chunk, model, processor, device):
    inputs = processor(chunk, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}
    with torch.no_grad():
        outputs = model(**inputs)
    output = outputs.reconstruction.data.squeeze().cpu().float().clamp_(0, 1).numpy()
    output = np.moveaxis(output, source=0, destination=-1)
    output_image = (output * 255.0).round().astype(np.uint8)
    return Image.fromarray(output_image)

@spaces.GPU(duration=120)
def main(image, model_choice, save_as_jpg=True):
    # Resize the input image
    image = resize_image(image)
    
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    model_paths = {
        "Pixel Perfect": "caidas/swin2SR-classical-sr-x4-64",
        "PSNR Match (Recommended)": "caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr"
    }
    
    processor = AutoImageProcessor.from_pretrained(model_paths[model_choice])
    model = Swin2SRForImageSuperResolution.from_pretrained(model_paths[model_choice]).to(device)
    
    # Split the image into chunks
    chunks = split_image(image)
    
    # Process each chunk
    upscaled_chunks = []
    for chunk, x, y in chunks:
        upscaled_chunk = upscale_chunk(chunk, model, processor, device)
        # Remove 32 pixels from bottom and right edges
        upscaled_chunk = upscaled_chunk.crop((0, 0, upscaled_chunk.width - 32, upscaled_chunk.height - 32))
        upscaled_chunks.append((upscaled_chunk, x * 4, y * 4))  # Multiply coordinates by 4 due to 4x upscaling
    
    # Stitch the chunks back together
    final_size = (image.width * 4 - 32, image.height * 4 - 32)  # Adjust for removed pixels
    upscaled_image = stitch_image(upscaled_chunks, final_size)
    
    if save_as_jpg:
        upscaled_image.save("upscaled_image.jpg", quality=95)
        return "upscaled_image.jpg"
    else:
        upscaled_image.save("upscaled_image.png")
        return "upscaled_image.png"

def gradio_interface(image, model_choice, save_as_jpg):
    try:
        result = main(image, model_choice, save_as_jpg)
        return result, None
    except Exception as e:
        return None, str(e)

interface = gr.Interface(
    fn=gradio_interface,
    inputs=[
        gr.Image(type="pil", label="Upload Image"),
        gr.Dropdown(
            choices=["PSNR Match (Recommended)", "Pixel Perfect"],
            label="Select Model",
            value="PSNR Match (Recommended)"
        ),
        gr.Checkbox(value=True, label="Save as JPEG"),
    ],
    outputs=[
        gr.File(label="Download Upscaled Image"),
        gr.Textbox(label="Error Message", visible=True)
    ],
    title="Image Upscaler",
    description="Upload an image, select a model, and upscale it. Images larger than 2048x2048 will be resized while maintaining aspect ratio. The image will be processed in 512x512 pixel chunks for efficient handling.",
)

interface.launch()