app.py

import spaces
import gradio as gr
import numpy as np
import torch

from chrislib.general import uninvert, invert, view, view_scale

from intrinsic.pipeline import load_models, run_pipeline

DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Instead of loading models at startup, we'll create a cache for models
model_cache = {}

def get_model(model_version):
    if model_version not in model_cache:
        model_cache[model_version] = load_models(model_version, device=DEVICE)
    return model_cache[model_version]

def generate_pipeline(models):
    def pipeline_func(image, **kwargs):
        return run_pipeline(models, image, **kwargs)
    
    return pipeline_func

@spaces.GPU
def process_image(image, model_version):
    # Check if image is provided
    if image is None:
        return [None, None, None]
        
    print(f"Processing with model version: {model_version}")
    print(image.shape)
    image = image.astype(np.single) / 255.

    # Get or load the selected model
    models = get_model(model_version)
    pipeline_func = generate_pipeline(models)
    
    result = pipeline_func(image, device=DEVICE, resize_conf=1024)
    
    return [view(result['hr_alb']), 1 - invert(result['dif_shd']), view_scale(result['pos_res'])]

with gr.Blocks(
    css="""
        #download {
            height: 118px;
        }
        .slider .inner {
            width: 5px;
            background: #FFF;
        }
        .viewport {
            aspect-ratio: 4/3;
        }
        .tabs button.selected {
            font-size: 20px !important;
            color: crimson !important;
        }
        h1 {
            text-align: center;
            display: block;
        }
        h2 {
            text-align: center;
            display: block;
        }
        h3 {
            text-align: center;
            display: block;
        }
        .md_feedback li {
            margin-bottom: 0px !important;
        }
        .image-gallery {
            display: flex;
            flex-wrap: wrap;
            gap: 10px;
            justify-content: center;
        }
        .image-gallery > * {
            flex: 1;
            min-width: 200px;
        }
        """,
) as demo:
    gr.Markdown(
            """
            # Colorful Diffuse Intrinsic Image Decomposition in the Wild
            <p align="center">
            <a title="Website" href="https://yaksoy.github.io/ColorfulShading/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
                <img src="https://www.obukhov.ai/img/badges/badge-website.svg">
            </a>
            <a title="Github" href="https://github.com/compphoto/Intrinsic" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
                <img src="https://img.shields.io/github/stars/compphoto/Intrinsic?label=GitHub%20%E2%98%85&logo=github&color=C8C" alt="badge-github-stars">
            </a>
        """
    )
    
    # Model version selector with information panel
    with gr.Row():
        model_version = gr.Dropdown(
            choices=["v2", "v2.1"], 
            value="v2", 
            label="Model Version",
            info="Select which model weights to use",
            scale=1
        )
        
        gr.Markdown("""
        The model may take a few seconds to load the first time you use it.
        Subsequent decompositions should be faster after the model is loaded.
        """)
    
    # Gallery-style layout for all images
    with gr.Row(elem_classes="image-gallery"):
        input_img = gr.Image(label="Input Image")
        alb_img = gr.Image(label="Albedo")
        shd_img = gr.Image(label="Diffuse Shading")
        dif_img = gr.Image(label="Diffuse Image")

    # Update to pass model_version to process_image
    input_img.change(
        process_image, 
        inputs=[input_img, model_version], 
        outputs=[alb_img, shd_img, dif_img]
    )
    # Add event handler for when model_version changes
    model_version.change(
        process_image,
        inputs=[input_img, model_version],
        outputs=[alb_img, shd_img, dif_img]
    )

demo.launch(show_error=True)