app.py

import gradio as gr
import spaces
from gradio_litmodel3d import LitModel3D

import os
import shutil
os.environ['SPCONV_ALGO'] = 'native'
from typing import *
import torch
import numpy as np
import imageio
from easydict import EasyDict as edict
from PIL import Image
from trellis.pipelines import TrellisVGGTTo3DPipeline
from trellis.representations import Gaussian, MeshExtractResult
from trellis.utils import render_utils, postprocessing_utils



MAX_SEED = np.iinfo(np.int32).max
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
os.makedirs(TMP_DIR, exist_ok=True)

def start_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    os.makedirs(user_dir, exist_ok=True)
    
    
def end_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    shutil.rmtree(user_dir)
def preprocess_image(image: Image.Image) -> Image.Image:
    """
    Preprocess the input image for 3D generation.
    
    This function is called when a user uploads an image or selects an example.
    It applies background removal and other preprocessing steps necessary for
    optimal 3D model generation.

    Args:
        image (Image.Image): The input image from the user

    Returns:
        Image.Image: The preprocessed image ready for 3D generation
    """
    processed_image = pipeline.preprocess_image(image)
    return processed_image


def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
    """
    Preprocess a list of input images for multi-image 3D generation.
    
    This function is called when users upload multiple images in the gallery.
    It processes each image to prepare them for the multi-image 3D generation pipeline.
    
    Args:
        images (List[Tuple[Image.Image, str]]): The input images from the gallery
        
    Returns:
        List[Image.Image]: The preprocessed images ready for 3D generation
    """
    images = [image[0] for image in images]
    processed_images = [pipeline.preprocess_image(image) for image in images]
    return processed_images


def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
    return {
        'gaussian': {
            **gs.init_params,
            '_xyz': gs._xyz.cpu().numpy(),
            '_features_dc': gs._features_dc.cpu().numpy(),
            '_scaling': gs._scaling.cpu().numpy(),
            '_rotation': gs._rotation.cpu().numpy(),
            '_opacity': gs._opacity.cpu().numpy(),
        },
        'mesh': {
            'vertices': mesh.vertices.cpu().numpy(),
            'faces': mesh.faces.cpu().numpy(),
        },
    }
    
    
def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
    gs = Gaussian(
        aabb=state['gaussian']['aabb'],
        sh_degree=state['gaussian']['sh_degree'],
        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],
        scaling_bias=state['gaussian']['scaling_bias'],
        opacity_bias=state['gaussian']['opacity_bias'],
        scaling_activation=state['gaussian']['scaling_activation'],
    )
    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda')
    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda')
    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda')
    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda')
    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda')
    
    mesh = edict(
        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),
        faces=torch.tensor(state['mesh']['faces'], device='cuda'),
    )
    
    return gs, mesh


def get_seed(randomize_seed: bool, seed: int) -> int:
    """
    Get the random seed for generation.
    
    This function is called by the generate button to determine whether to use
    a random seed or the user-specified seed value.
    
    Args:
        randomize_seed (bool): Whether to generate a random seed
        seed (int): The user-specified seed value
        
    Returns:
        int: The seed to use for generation
    """
    return np.random.randint(0, MAX_SEED) if randomize_seed else seed


@spaces.GPU
def generate_and_extract_glb(
    multiimages: List[Tuple[Image.Image, str]],
    seed: int,
    ss_guidance_strength: float,
    ss_sampling_steps: int,
    slat_guidance_strength: float,
    slat_sampling_steps: int,
    multiimage_algo: Literal["multidiffusion", "stochastic"],
    mesh_simplify: float,
    texture_size: int,
    req: gr.Request,
) -> Tuple[dict, str, str, str]:
    """
    Convert an image to a 3D model and extract GLB file.

    Args:
        image (Image.Image): The input image.
        multiimages (List[Tuple[Image.Image, str]]): The input images in multi-image mode.
        is_multiimage (bool): Whether is in multi-image mode.
        seed (int): The random seed.
        ss_guidance_strength (float): The guidance strength for sparse structure generation.
        ss_sampling_steps (int): The number of sampling steps for sparse structure generation.
        slat_guidance_strength (float): The guidance strength for structured latent generation.
        slat_sampling_steps (int): The number of sampling steps for structured latent generation.
        multiimage_algo (Literal["multidiffusion", "stochastic"]): The algorithm for multi-image generation.
        mesh_simplify (float): The mesh simplification factor.
        texture_size (int): The texture resolution.

    Returns:
        dict: The information of the generated 3D model.
        str: The path to the video of the 3D model.
        str: The path to the extracted GLB file.
        str: The path to the extracted GLB file (for download).
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    image_files = [image[0] for image in multiimages]

    # Generate 3D model
    outputs = pipeline.run(
        image=image_files,
        seed=seed,
        formats=["gaussian", "mesh"],
        preprocess_image=False,
        sparse_structure_sampler_params={
            "steps": ss_sampling_steps,
            "cfg_strength": ss_guidance_strength,
        },
        slat_sampler_params={
            "steps": slat_sampling_steps,
            "cfg_strength": slat_guidance_strength,
        },
        mode=multiimage_algo,
    )

    # Render video
    video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
    video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
    video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
    video_path = os.path.join(user_dir, 'sample.mp4')
    imageio.mimsave(video_path, video, fps=15)
    
    # Extract GLB
    gs = outputs['gaussian'][0]
    mesh = outputs['mesh'][0]
    glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
    glb_path = os.path.join(user_dir, 'sample.glb')
    glb.export(glb_path)
    
    # Pack state for optional Gaussian extraction
    state = pack_state(gs, mesh)
    
    torch.cuda.empty_cache()
    return state, video_path, glb_path, glb_path


@spaces.GPU
def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
    """
    Extract a Gaussian splatting file from the generated 3D model.
    
    This function is called when the user clicks "Extract Gaussian" button.
    It converts the 3D model state into a .ply file format containing
    Gaussian splatting data for advanced 3D applications.

    Args:
        state (dict): The state of the generated 3D model containing Gaussian data
        req (gr.Request): Gradio request object for session management

    Returns:
        Tuple[str, str]: Paths to the extracted Gaussian file (for display and download)
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    gs, _ = unpack_state(state)
    gaussian_path = os.path.join(user_dir, 'sample.ply')
    gs.save_ply(gaussian_path)
    torch.cuda.empty_cache()
    return gaussian_path, gaussian_path


def prepare_multi_example() -> List[Image.Image]:
    multi_case = list(set([i.split('_')[0] for i in os.listdir("assets/example_multi_image")]))
    images = []
    for case in multi_case:
        _images = []
        for i in range(1, 9):
            if os.path.exists(f'assets/example_multi_image/{case}_{i}.png'):
                img = Image.open(f'assets/example_multi_image/{case}_{i}.png')
                W, H = img.size
                img = img.resize((int(W / H * 512), 512))
                _images.append(np.array(img))
        if len(_images) > 0:
            images.append(Image.fromarray(np.concatenate(_images, axis=1)))
    return images


def split_image(image: Image.Image) -> List[Image.Image]:
    """
    Split a multi-view image into separate view images.
    
    This function is called when users select multi-image examples that contain
    multiple views in a single concatenated image. It automatically splits them
    based on alpha channel boundaries and preprocesses each view.
    
    Args:
        image (Image.Image): A concatenated image containing multiple views
        
    Returns:
        List[Image.Image]: List of individual preprocessed view images
    """
    image = np.array(image)
    alpha = image[..., 3]
    alpha = np.any(alpha>0, axis=0)
    start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()
    end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()
    images = []
    for s, e in zip(start_pos, end_pos):
        images.append(Image.fromarray(image[:, s:e+1]))
    return [preprocess_image(image) for image in images]


with gr.Blocks(delete_cache=(600, 600)) as demo:
    gr.Markdown("""
    ## Multi-view images to 3D Asset with [ReconViaGen](https://jiahao620.github.io/reconviagen/)
    * Upload an image and click "Generate & Extract GLB" to create a 3D asset and automatically extract the GLB file.
    * If you want the Gaussian file as well, click "Extract Gaussian" after generation.
    * If the image has alpha channel, it will be used as the mask. Otherwise, we use `rembg` to remove the background.
    
    ✨This demo is partial. We will release the whole model later. Stay tuned!✨
    """)
    
    with gr.Row():
        with gr.Column():
            with gr.Tabs() as input_tabs:
                with gr.Tab(label="Multiple Images", id=0) as multiimage_input_tab:
                    image_prompt = gr.Image(label="Image Prompt", format="png", visible=False, image_mode="RGBA", type="pil", height=300)
                    multiimage_prompt = gr.Gallery(label="Image Prompt", format="png", type="pil", height=300, columns=3)
                    gr.Markdown("""
                        Input different views of the object in separate images. 
                    """)
        
            with gr.Accordion(label="Generation Settings", open=False):
                seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=False)
                gr.Markdown("Stage 1: Sparse Structure Generation")
                with gr.Row():
                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=30, step=1)
                gr.Markdown("Stage 2: Structured Latent Generation")
                with gr.Row():
                    slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
                    slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
                multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], label="Multi-image Algorithm", value="multidiffusion")
            
            with gr.Accordion(label="GLB Extraction Settings", open=False):
                mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
                texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)

            generate_btn = gr.Button("Generate & Extract GLB", variant="primary")
            extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
            gr.Markdown("""
                        *NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
                        """)

        with gr.Column():
            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
            model_output = LitModel3D(label="Extracted GLB/Gaussian", exposure=10.0, height=300)
            
            with gr.Row():
                download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
                download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)  
    
    output_buf = gr.State()

    # Example images at the bottom of the page
    with gr.Row() as multiimage_example:
        examples_multi = gr.Examples(
            examples=prepare_multi_example(),
            inputs=[image_prompt],
            fn=split_image,
            outputs=[multiimage_prompt],
            run_on_click=True,
            examples_per_page=8,
        )

    # Handlers
    demo.load(start_session)
    demo.unload(end_session)


    multiimage_prompt.upload(
        preprocess_images,
        inputs=[multiimage_prompt],
        outputs=[multiimage_prompt],
    )

    generate_btn.click(
        get_seed,
        inputs=[randomize_seed, seed],
        outputs=[seed],
    ).then(
    #     lambda: [None, None, None, None],  # 先清空 video_output
    #     inputs=[],
    #     outputs=[video_output, model_output, download_glb, download_gs],
    # ).then(
        generate_and_extract_glb,
        inputs=[multiimage_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo, mesh_simplify, texture_size],
        outputs=[output_buf, video_output, model_output, download_glb],
    ).then(
        lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),
        outputs=[extract_gs_btn, download_glb],
    )

    video_output.clear(
        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False), gr.Button(interactive=False)]),
        outputs=[extract_gs_btn, download_glb, download_gs],
    )
    
    extract_gs_btn.click(
        extract_gaussian,
        inputs=[output_buf],
        outputs=[model_output, download_gs],
    ).then(
        lambda: gr.Button(interactive=True),
        outputs=[download_gs],
    )

    model_output.clear(
        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),
        outputs=[download_glb, download_gs],
    )
    

# Launch the Gradio app
if __name__ == "__main__":
    pipeline = TrellisVGGTTo3DPipeline.from_pretrained("Stable-X/trellis-vggt-v0-1")
    pipeline.cuda()
    pipeline.VGGT_model.cuda()
    pipeline.birefnet_model.cuda()
    try:
        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))    # Preload rembg
    except:
        pass
    demo.launch(share=True)