app.py

import gradio as gr
from PIL import Image
import os
import torch
from models import get_available_models
from inference import run_inference

device = "cuda" if torch.cuda.is_available() else "cpu"
models_dict = get_available_models(model_dir="models", device=device)

def save_image(img, prefix):
    if img is None:
        return None
    path = f"{prefix}_output.png"
    img.save(path)
    return path

def process_image(img, use_x8):
    if img is None:
        return [None]*9 # 3 sliders, 3 standalone images, 3 file paths
    
    # Run SR
    results = run_inference(img, models_dict, x8_mode=use_x8, device=device)
    
    # Original image scaled to match the resolution for slider comparison
    scale = 8 if use_x8 else 4
    w, h = img.size
    img_scaled = img.resize((w * scale, h * scale), Image.BICUBIC)
    
    outputs = []
    image_outputs = []
    file_outputs = []
    
    for m in ["srcnn", "satlas", "esrgan"]:
        res_img = results[m]
        if res_img is not None:
            outputs.append((img_scaled, res_img))
            image_outputs.append(res_img)
            file_outputs.append(save_image(res_img, m))
        else:
            outputs.append(None)
            image_outputs.append(None)
            file_outputs.append(None)
            
    return (*outputs, *image_outputs, *file_outputs)

css = """
.model-panel { border: 1px solid #ccc; padding: 10px; border-radius: 8px; margin-bottom: 20px; }
"""

with gr.Blocks(title="Super Resolution Comparative Space") as demo:
    gr.Markdown("# High-Res Super Resolution with Tile-Based Inference")
    gr.Markdown("Compare SRCNN, SpectraGAN, and ESRGAN on a single image.")
    
    with gr.Row():
        with gr.Column(scale=1):
            input_image = gr.Image(type="pil", label="Input Image")
            x8_checkbox = gr.Checkbox(label="Enable x8 Post-Resize", info="bicubic x2 pass on top of the x4 result")
            submit_btn = gr.Button("Upscale", variant="primary")
            
        with gr.Column(scale=3):
            # Model panels
            sliders = {}
            images = {}
            files = {}
            names = ["srcnn", "satlas", "esrgan"]
            disp_names = ["SRCNN x4", "SpectraGAN x4", "ESRGAN x4"]
            for i, name in enumerate(names):
                with gr.Group(elem_classes="model-panel"):
                    gr.Markdown(f"### {disp_names[i]}")
                    from gradio_imageslider import ImageSlider
                    sliders[name] = ImageSlider(label=f"{disp_names[i]} Comparison", show_label=False)
                    images[name] = gr.Image(type="pil", label="Final Output", interactive=False)
                    files[name] = gr.File(label=f"Download {disp_names[i]} PNG", interactive=False)
    
    # Collect outputs in order: 3 sliders, then 3 images, then 3 files
    outputs = [sliders["srcnn"], sliders["satlas"], sliders["esrgan"],
               images["srcnn"], images["satlas"], images["esrgan"],
               files["srcnn"], files["satlas"], files["esrgan"]]
               
    submit_btn.click(
        fn=process_image,
        inputs=[input_image, x8_checkbox],
        outputs=outputs
    )

if __name__ == "__main__":
    demo.launch(css=css, server_name="0.0.0.0", server_port=7860, ssr_mode=False, share=False)