app.py

import torch
import gradio as gr
from shap_e.models.download import load_model
from shap_e.diffusion.sample import sample_latents
from shap_e.util.notebooks import decode_latent_mesh
from PIL import Image

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

# Load Shap-E models
print("Loading Shap-E models (this may take a bit)...")
transmitter_model = load_model("transmitter", device=device)
image_model = load_model("image300M", device=device)

def generate_3d(image: Image.Image):
    """Takes an uploaded image and returns path to generated 3D model (.obj)"""
    img = image.convert("RGB")

    # Sample latents (updated for latest Shap-E API)
    latents = sample_latents(
        batch_size=1,
        model=image_model,
        model_kwargs=dict(images=[img]),
        diffusion=None,
        clip_denoised=True,
        use_fp16=False,
        use_karras=False,
        karras_steps=64,
        sigma_min=0.002,
        sigma_max=80,
        s_churn=0.0,
        guidance_scale=3.0,
        device=device
    )

    # Decode into mesh
    mesh = decode_latent_mesh(transmitter_model, latents[0])

    # Save output
    output_path = "output.obj"
    with open(output_path, "w") as f:
        mesh.write_obj(f)

    return output_path

# Gradio interface
demo = gr.Interface(
    fn=generate_3d,
    inputs=gr.Image(type="pil"),
    outputs=gr.File(file_types=[".obj"]),
    title="Shap-E: 2D → 3D Model",
    description="Upload a 2D image and download a generated 3D model (.obj)"
)

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