app.py

import os
import time
import gradio as gr
import torch
from PIL import Image
import numpy as np
from torchvision import transforms
import tempfile

# Configuration
CHUNK_SIZE = 256
SCALE_FACTOR = 4
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'

# Load model
model = torch.jit.load('4xTextures_GTAV_rgt-s.pth', map_location=DEVICE)
model.eval()

def process_chunk(chunk):
    """Process a single chunk through the model"""
    preprocess = transforms.Compose([
        transforms.ToTensor()
    ])
    
    img_tensor = preprocess(chunk).unsqueeze(0).to(DEVICE)
    
    with torch.no_grad():
        output = model(img_tensor)
    
    output = output.squeeze().cpu().clamp(0, 1).numpy()
    return Image.fromarray((output * 255).astype(np.uint8))

def split_image(img, chunk_size):
    """Split image into chunks"""
    width, height = img.size
    chunks = []
    positions = []
    
    for y in range(0, height, chunk_size):
        for x in range(0, width, chunk_size):
            box = (x, y, x+chunk_size, y+chunk_size)
            chunks.append(img.crop(box))
            positions.append((x, y))
    
    return chunks, positions

def merge_chunks(chunk_files, positions, target_size):
    """Merge processed chunks into final image"""
    merged_img = Image.new('RGB', target_size)
    for (x, y), path in zip(positions, chunk_files):
        chunk = Image.open(path)
        merged_img.paste(chunk, (x*SCALE_FACTOR, y*SCALE_FACTOR))
        os.remove(path)  # Cleanup immediately
    return merged_img

def upscale_image(input_img):
    """Main processing function"""
    start_time = time.time()
    
    # Validate input
    if input_img.size[0] != input_img.size[1]:
        raise ValueError("Input image must be square")
    
    original_size = input_img.size[0]
    target_size = original_size * SCALE_FACTOR
    
    # Split into chunks
    chunks, positions = split_image(input_img, CHUNK_SIZE)
    total_chunks = len(chunks)
    
    # Create temporary directory
    temp_dir = tempfile.mkdtemp()
    chunk_files = []
    
    # Process chunks
    for i, (chunk, (x, y)) in enumerate(zip(chunks, positions)):
        chunk_start = time.time()
        
        # Process chunk
        upscaled = process_chunk(chunk)
        
        # Save to temp file
        chunk_path = os.path.join(temp_dir, f'chunk_{x}_{y}.png')
        upscaled.save(chunk_path)
        chunk_files.append(chunk_path)
        
        # Calculate progress
        elapsed = time.time() - chunk_start
        progress = (i + 1) / total_chunks * 100
        print(f"Processed chunk {i+1}/{total_chunks} ({progress:.1f}%) - {elapsed:.2f}s")
    
    # Merge chunks
    final_image = merge_chunks(chunk_files, positions, (target_size, target_size))
    os.rmdir(temp_dir)  # Cleanup temp directory
    
    total_time = time.time() - start_time
    print(f"Total processing time: {total_time:.2f}s")
    return final_image

# Gradio interface
demo = gr.Interface(
    fn=upscale_image,
    inputs=gr.Image(type="pil", label="Input Image"),
    outputs=gr.Image(type="pil", label="Upscaled Image"),
    title="4x Texture Upscaler (Low Memory)",
    description="Upscale large square textures using RGT model. Accepts 256, 512, 1024, 2048, 4096, 8192px images.",
    allow_flagging="never"
)

if __name__ == "__main__":
    demo.launch()