app.py

import torch
import torchvision
import numpy as np

import os
from omegaconf import OmegaConf
from PIL import Image 
import spaces 

from utils.app_utils import (
    remove_background, 
    resize_foreground, 
    set_white_background,
    resize_to_128,
    to_tensor,
    get_source_camera_v2w_rmo_and_quats,
    get_target_cameras,
    export_to_obj)

import imageio

from scene.gaussian_predictor import GaussianSplatPredictor

import gradio as gr

import rembg

from huggingface_hub import hf_hub_download

def main():

    if torch.cuda.is_available():
        device = "cuda:0"
    else:
        device = "cpu"

    model_cfg = OmegaConf.load(
        os.path.join(
            os.path.dirname(os.path.abspath(__file__)), 
                    "config.yaml"
                    ))

    model_path = hf_hub_download(repo_id="szymanowiczs/splatter-image-multi-category-v1", 
                                 filename="model_latest.pth")

    
    model = GaussianSplatPredictor(model_cfg)

    ckpt_loaded = torch.load(model_path, map_location="cpu")
    model.load_state_dict(ckpt_loaded["model_state_dict"])
    model.to(device)

    # ============= image preprocessing =============
    rembg_session = rembg.new_session()

    def check_input_image(input_image):
        if input_image is None:
            raise gr.Error("No image uploaded!")

    def preprocess(input_image, preprocess_background=True, foreground_ratio=0.65):
        # 0.7 seems to be a reasonable foreground ratio
        if preprocess_background:
            image = input_image.convert("RGB")
            image = remove_background(image, rembg_session)
            image = resize_foreground(image, foreground_ratio)
            image = set_white_background(image)
        else:
            image = input_image
            if image.mode == "RGBA":
                image = set_white_background(image)
        image = resize_to_128(image)
        return image

    ply_out_path = f'./mesh.ply'

    @spaces.GPU()
    def reconstruct_and_export(image):
        """
        Passes image through model, outputs reconstruction in form of a dict of tensors.
        """
        image = to_tensor(image).to(device)
        view_to_world_source, rot_transform_quats = get_source_camera_v2w_rmo_and_quats()
        view_to_world_source = view_to_world_source.to(device)
        rot_transform_quats = rot_transform_quats.to(device)

        reconstruction_unactivated = model(
            image.unsqueeze(0).unsqueeze(0),
            view_to_world_source,
            rot_transform_quats,
            None,
            activate_output=False)

        # export reconstruction to ply
        export_to_obj(reconstruction_unactivated, ply_out_path)

        return ply_out_path

    with gr.Blocks() as demo:
        gr.Markdown(
            """

            # Splatter Image Demo
            [Splatter Image](https://github.com/szymanowiczs/splatter-image) (CVPR 2024) is a fast, super cheap to train method for object 3D reconstruction from a single image. 
            The model used in the demo was trained on **Objaverse-LVIS on 2 A6000 GPUs for 3.5 days**.
            On NVIDIA V100 GPU, reconstruction can be done at 38FPS and rendering at 588FPS.
            Upload an image of an object to see how the Splatter Image does.
            
            **Comments:**
            1. The first example you upload should take about 4.5 seconds (with preprocessing, saving and overhead), the following take about 1.5s.
            2. The model does not work well on photos of humans.
            3. The 3D viewer shows a .ply mesh extracted from a mix of 3D Gaussians. Artefacts might show - see video for more faithful results.
            4. Best results are achieved on the datasets described in the [repository](https://github.com/szymanowiczs/splatter-image) using that code. This demo is experimental.
            5. Our model might not be better than some state-of-the-art methods, but it is of comparable quality and is **much** cheaper to train and run.
            """
            )
        with gr.Row(variant="panel"):
            with gr.Column():
                with gr.Row():
                    input_image = gr.Image(
                        label="Input Image",
                        image_mode="RGBA",
                        sources="upload",
                        type="pil",
                        elem_id="content_image",
                    )
                    processed_image = gr.Image(label="Processed Image", interactive=False)
                with gr.Row():
                    with gr.Group():
                        preprocess_background = gr.Checkbox(
                            label="Remove Background", value=True
                        )
                with gr.Row():
                    submit = gr.Button("Generate", elem_id="generate", variant="primary")
            with gr.Column():
                with gr.Row():
                    with gr.Tab("Reconstruction"):
                        output_model = gr.Model3D(
                            height=512,
                            label="Output Model",
                            interactive=False
                        )

        submit.click(fn=check_input_image, inputs=[input_image]).success(
            fn=preprocess,
            inputs=[input_image, preprocess_background],
            outputs=[processed_image],
        ).success(
            fn=reconstruct_and_export,
            inputs=[processed_image],
            outputs=[output_model],
        )

    demo.queue(max_size=1)
    demo.launch()


if __name__ == "__main__":
    main()

# gradio app interface