app.py

import gradio as gr
import pyiqa
import torch
import pandas as pd
import numpy as np
from PIL import Image

# Initialize device
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

# Load Models (Pre-selected for variety: Statistical, Perception, Deep Learning)
# Note: You can add more like 'niqe', 'clipiqa', etc.
metrics = {
    "BRISQUE (Lower=Better)": pyiqa.create_metric('brisque', device=device),
    "MUSIQ (Higher=Better)": pyiqa.create_metric('musiq', device=device),
    "NIMA (Higher=Better)": pyiqa.create_metric('nima', device=device)
}

def analyze_image(input_img):
    if input_img is None:
        return None
    
    np_img = np.array(input_img)
    np_img = np_img / 255.0
    # Standardize input for pyiqa
    img_tensor = pyiqa.utils.img2tensor(np_img).unsqueeze(0).to(device)
    
    results = []
    with torch.no_grad():
        for name, metric in metrics.items():
            score = metric(img_tensor).item()
            results.append({"Model": name, "Score": round(score, 4)})
    
    return pd.DataFrame(results)

# Create Gradio Interface
with gr.Blocks(title="Multi-Model IQA Comparison") as demo:
    gr.Markdown("# 📸 Image Quality Checker")
    gr.Markdown("Upload an image to evaluate its quality score across various SOTA models.")
    
    with gr.Row():
        image_input = gr.Image(type="pil", label="Input Image")
        result_table = gr.Dataframe(label="Model Scores")
        
    btn = gr.Button("Get Scores")
    btn.click(fn=analyze_image, inputs=image_input, outputs=result_table)

    gr.Markdown("### Model Descriptions:")
    gr.Markdown("- **BRISQUE:** Statistical model focusing on naturalness and noise.")
    gr.Markdown("- **MUSIQ:** Transformer-based model optimized for varied image sizes.")
    gr.Markdown("- **NIMA:** Neural Aesthetic model that mimics human perception.")

demo.launch()