app.py

import gradio as gr
import torch
import spaces
from omegaconf import OmegaConf
import torch
import torchvision.transforms.functional as TF
from idf.models.lit_a_denoising import LitADenoising

# --- Model Loading ---
# Load Pretrained IDF Model
config = OmegaConf.load("configs/models/idfnet.yaml") 
device = "cuda" if torch.cuda.is_available() else "cpu" 
model_pth = "pretrained_models/idf_g_15.ckpt" 

# Load the model
model = LitADenoising.load_from_checkpoint(model_pth, strict=True, map_location=device, 
                                           denoiser_config=config.params.denoiser_config, 
                                           loss_config=config.params.loss_config)
model.to(device) 
model.eval().freeze()

# --- Main Inference Function  ---
@spaces.GPU()
def infer(
    img, 
    adaptive_iter, 
    max_iter = None,
):
    if img is None:
        return None
    x = TF.to_tensor(img).unsqueeze(0).to(device)
    x = model(x, adaptive_iter, max_iter)
    x = torch.clamp(x, 0.0, 1.0)
    
    return TF.to_pil_image(x.squeeze(0))

# --- Examples and UI Layout ---
examples = []

css = """
#col-container {
    margin: 0 auto;
    max-width: 1024px;
}
#edit_text{
    margin-top: -62px !important
}
.badge-row {
    display: flex;
    justify-content: center;
    align-items: center;
    gap: 12px;
    flex-wrap: nowrap; /* force single line; change to wrap if needed */
    margin-top: 8px;
}
.badge-row a { text-decoration: none; }
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.HTML(
            """
            <div align='center'>
                <img src="https://dongjinkim9.github.io/projects/idf/assets/idf_logo.png" alt="IDF Logo" width="200" style="display: block; margin: 0 auto;">
                <h1>IDF: <span style="color: red;">I</span>terative <span style="color: red;">D</span>ynamic <span style="color: red;">F</span>iltering Networks <br>for Generalizable Image Denoising</h1>
                <div class="badge-row">
                    <a href="https://arxiv.org/abs/2508.19649" target="_blank">
                        <img src="https://img.shields.io/badge/Arxiv-📄Paper-8A2BE2" alt="Arxiv" />
                    </a>
                    <a href="https://dongjinkim9.github.io/projects/idf" target="_blank">
                        <img src="https://img.shields.io/badge/Project-📖Page-8A2BE2" alt="Project Page" />
                    </a>
                    <a href="https://github.com/dongjinkim9/IDF" target="_blank">
                        <img src="https://img.shields.io/badge/Github-💻Code-8A2BE2" alt="Github Code" />
                    </a>
                </div>
            </div>

            <br>
            <blockquote>
            <p><i>Even though IDF is trained with extremely limited data (e.g., a single-level Gaussian noise), 
            it generalizes effectively to diverse unseen noise types and levels with only ~0.04M parameters.</i></p>
            </blockquote>
            
            """
        ) 

        with gr.Row():
            with gr.Column():
                input_image = gr.Image(label="Noisy Image", show_label=True, type="pil")

            result = gr.Image(label="Denoised Image", show_label=True, type="pil")
        with gr.Row():
            clear_button = gr.Button("Clear", variant="secondary")
            run_button = gr.Button("Denoise!", variant="primary")

        with gr.Accordion("Advanced Settings", open=False):
            adaptive_iter = gr.Checkbox(value=False, label="Dynamic Iteration Control")
            with gr.Row():
                max_iter = gr.Slider(minimum=1, 
                                     maximum=20, 
                                     value=10, 
                                     step=1, 
                                     label="Maximum Denoising Iteration")



        gr.Examples(
            examples=[
                ["assets/demo/noisy/DSLR.png", False, 10],
                ["assets/demo/noisy/Smartphone.png", False, 10],
                ["assets/demo/noisy/Salt_Pepper.png", False, 10],
                ["assets/demo/noisy/Gaussian.png", False, 10],
                ["assets/demo/noisy/Spatial_Gaussian.png", False, 10],
                ["assets/demo/noisy/Poisson.png", False, 10],
                ["assets/demo/noisy/Speckle.png", False, 10],
                ["assets/demo/noisy/Mixture.png", False, 10]
            ],
            inputs=[
                input_image, 
                adaptive_iter, 
                max_iter
            ],
            outputs=[result],
            fn=infer,
            cache_examples=True,
            cache_mode='lazy',
            label="Examples (Noise Types)",
            example_labels=["📸DSLR","📱Smartphone","🧂Salt & Pepper","🌫️Gaussian","🌀Spatial Gaussian","🎲Poisson","✨Speckle", "🧪Mixture"],
        )

    # Define the event triggers for the run button.
    run_button.click(
        fn=infer,
        inputs=[
            input_image, 
            adaptive_iter, 
            max_iter
        ],
        outputs=[result],
    )

    # Define the clear button functionality.
    clear_button.click(
        fn=lambda: [None, False, 10],
        outputs=[input_image, adaptive_iter, max_iter]
    )

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