app.py

import torch

import os
from omegaconf import OmegaConf
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,
    export_to_obj)


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_path = hf_hub_download(repo_id="szymanowiczs/splatter-image-v1", 
                                 filename="config_objaverse.yaml")
    model_path = hf_hub_download(repo_id="szymanowiczs/splatter-image-v1", 
                                 filename="model_latest.pth")

    model_cfg = OmegaConf.load(model_cfg_path)    
    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

    css = """
        h1 {
            text-align: center;
            display:block;
        }
        """

    with gr.Blocks(css=css) as demo:
        gr.Markdown(
            """
            # Splatter Image

            **Splatter Image (CVPR 2024)** [[code](https://github.com/szymanowiczs/splatter-image), [project page](https://szymanowiczs.github.io/splatter-image)] 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**.
            Locally, on an NVIDIA V100 GPU, reconstruction (forward pass of the network) can be done at 38FPS and rendering (with Gaussian Splatting) at 588FPS.
            Upload an image of an object or click on one of the provided examples to see how the Splatter Image does.
            The 3D viewer will render a .ply object exported from the 3D Gaussians, which is only an approximation.
            For best results run the demo locally and render locally with Gaussian Splatting - to do so, clone the [main repository](https://github.com/szymanowiczs/splatter-image).
            """
            )
        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.Row(variant="panel"): 
                    gr.Examples(
                        examples=[
                            './demo_examples/01_bigmac.png',
                            './demo_examples/02_hydrant.jpg',
                            './demo_examples/03_spyro.png',
                            './demo_examples/04_lysol.png',
                            './demo_examples/05_pinapple_bottle.png',
                            './demo_examples/06_unsplash_broccoli.png',
                            './demo_examples/07_objaverse_backpack.png',
                            './demo_examples/08_unsplash_chocolatecake.png',
                            './demo_examples/09_realfusion_cherry.png',
                            './demo_examples/10_triposr_teapot.png'
                        ],
                        inputs=[input_image],
                        cache_examples=False,
                        label="Examples",
                        examples_per_page=20,
                    )
            with gr.Column():
                with gr.Row():
                    with gr.Tab("Reconstruction"):
                        output_model = gr.Model3D(
                            height=512,
                            label="Output Model",
                            interactive=False
                        )

        gr.Markdown(
        """
            ## Comments:
            1. If you run the demo online, the first example you upload should take about 4.5 seconds (with preprocessing, saving and overhead), the following take about 1.5s.
            2. The 3D viewer shows a .ply mesh extracted from a mix of 3D Gaussians. This is only an approximations and artefacts might show.
            3. Known limitations include:
            - a black dot appearing on the model from some viewpoints
            - see-through parts of objects, especially on the back: this is due to the model performing less well on more complicated shapes
            - back of objects are blurry: this is a model limiation due to it being deterministic
            4. Our model is of comparable quality to state-of-the-art methods, and is **much** cheaper to train and run.

            ## How does it work?

            Splatter Image formulates 3D reconstruction as an image-to-image translation task. It maps the input image to another image, 
            in which every pixel represents one 3D Gaussian and the channels of the output represent parameters of these Gaussians, including their shapes, colours and locations.
            The resulting image thus represents a set of Gaussians (almost like a point cloud) which reconstruct the shape and colour of the object.
            The method is very cheap: the reconstruction amounts to a single forward pass of a neural network with only 2D operators (2D convolutions and attention).
            The rendering is also very fast, due to using Gaussian Splatting.
            Combined, this results in very cheap training and high-quality results.
            For more results see the [project page](https://szymanowiczs.github.io/splatter-image) and the [CVPR article](https://arxiv.org/abs/2312.13150).
            """
        )



        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