app.py

import os
os.environ.setdefault("SPCONV_ALGO", "native")
os.environ.setdefault("ATTN_BACKEND", "xformers")
os.environ.setdefault("SPARSE_ATTN_BACKEND", "xformers")

import subprocess
import sys

# Install gradio_litmodel3d ignoring its over-restrictive gradio<5 cap.
try:
    import gradio_litmodel3d  # noqa: F401
except ImportError:
    subprocess.check_call(
        [sys.executable, "-m", "pip", "install", "--no-deps", "gradio_litmodel3d==0.0.1"],
    )

import spaces
import torch
import ctypes
import tempfile

CUDA_HOME = "/cuda-image/usr/local/cuda-13.0"
CUDA_LIBDIR = os.path.join(CUDA_HOME, "lib64")


@spaces.GPU(duration=600)
def _first_gpu_setup():
    need = {}
    for name, modname in [
        ("nvdiffrast", "nvdiffrast"),
        ("diff_gaussian_rasterization", "diff_gaussian_rasterization"),
    ]:
        try:
            __import__(modname)
        except ImportError:
            need[name] = True
    if not need:
        return

    patch_dir = tempfile.mkdtemp(prefix="torch_cuda_patch_")
    with open(os.path.join(patch_dir, "sitecustomize.py"), "w") as f:
        f.write(
            "try:\n"
            "    import torch.utils.cpp_extension as _c\n"
            "    _c._check_cuda_version = lambda *a, **k: None\n"
            "except Exception:\n"
            "    pass\n"
        )

    env = os.environ.copy()
    env["CUDA_HOME"] = CUDA_HOME
    env["CUDA_PATH"] = CUDA_HOME
    env["PATH"] = os.path.join(CUDA_HOME, "bin") + os.pathsep + env.get("PATH", "")
    env["PYTHONPATH"] = patch_dir + os.pathsep + env.get("PYTHONPATH", "")
    env["TORCH_CUDA_ARCH_LIST"] = "12.0"

    subprocess.check_call(
        [sys.executable, "-m", "pip", "install", "--no-deps",
         "setuptools", "wheel", "ninja", "packaging"],
    )

    if "nvdiffrast" in need:
        subprocess.check_call(
            [sys.executable, "-m", "pip", "install",
             "--no-build-isolation", "--no-deps",
             "git+https://github.com/NVlabs/nvdiffrast/"],
            env=env,
        )
    if "diff_gaussian_rasterization" in need:
        mip = tempfile.mkdtemp(prefix="mip_")
        subprocess.check_call(
            ["git", "clone", "--recursive", "--depth=1",
             "https://github.com/autonomousvision/mip-splatting.git", mip],
        )
        subprocess.check_call(
            [sys.executable, "-m", "pip", "install",
             "--no-build-isolation", "--no-deps",
             os.path.join(mip, "submodules", "diff-gaussian-rasterization")],
            env=env,
        )


_first_gpu_setup()
try:
    ctypes.CDLL(os.path.join(CUDA_LIBDIR, "libcudart.so.13"), mode=ctypes.RTLD_GLOBAL)
    os.environ["LD_LIBRARY_PATH"] = CUDA_LIBDIR + os.pathsep + os.environ.get("LD_LIBRARY_PATH", "")
except OSError:
    pass

# xformers on the Blackwell (sm_120) ZeroGPU container is built without CUDA
# extensions for any FwOp: cutlassF-pt rejects compute capability >= (9, 0)
# ("too new") and FlashAttn3 is Hopper-only. Reroute xformers.ops.memory_efficient_attention
# (used by DINOv2, VGGT, trellis dense+sparse paths) to torch.nn.functional.scaled_dot_product_attention,
# which is CUDA-native on torch 2.10/2.11 and supports sm_120. Must be patched BEFORE
# anything that calls memory_efficient_attention is imported.
try:
    import xformers.ops as _xops
    import torch.nn.functional as _F
    from xformers.ops.fmha.attn_bias import BlockDiagonalMask as _BlockDiagonalMask

    def _bdm_starts(seqinfo):
        # xformers' BlockDiagonalMask sub-attribute. Try the public python-list view first;
        # otherwise pull from the tensor and tolist().
        for attr in ("seqstart_py", "_seqstart_py"):
            v = getattr(seqinfo, attr, None)
            if v is not None:
                return list(v)
        t = getattr(seqinfo, "seqstart", None)
        if t is not None:
            return t.detach().cpu().tolist()
        raise AttributeError("BlockDiagonalMask seqinfo has no seqstart_py / seqstart")

    def _mea_sdpa(q, k, v, attn_bias=None, p=0.0, scale=None, op=None):
        # q, k, v shapes: [B, M, H, K] (xformers convention). SDPA wants [B, H, M, K].
        if isinstance(attn_bias, _BlockDiagonalMask):
            # Block-diagonal mask used by trellis sparse attention to batch
            # variable-length sequences in a single dense tensor. Materialize each
            # block separately and concat — SDPA has no block-diagonal kernel.
            q_starts = _bdm_starts(attn_bias.q_seqinfo)
            k_starts = _bdm_starts(attn_bias.k_seqinfo)
            outs = []
            # q,k,v come in as [1, total_tokens, H, K]
            for i in range(len(q_starts) - 1):
                qs, qe = q_starts[i], q_starts[i + 1]
                ks, ke = k_starts[i], k_starts[i + 1]
                qi = q[:, qs:qe].transpose(1, 2)  # [1, H, Lq, K]
                ki = k[:, ks:ke].transpose(1, 2)  # [1, H, Lk, K]
                vi = v[:, ks:ke].transpose(1, 2)
                oi = _F.scaled_dot_product_attention(qi, ki, vi, dropout_p=p, scale=scale)
                outs.append(oi.transpose(1, 2))   # back to [1, Li, H, K]
            return torch.cat(outs, dim=1)

        attn_mask = None
        if attn_bias is not None and hasattr(attn_bias, "materialize"):
            attn_mask = attn_bias.materialize((q.shape[0], q.shape[2], q.shape[1], k.shape[1]),
                                              dtype=q.dtype, device=q.device)
        elif attn_bias is not None:
            attn_mask = attn_bias

        qh = q.transpose(1, 2)  # [B, H, M, K]
        kh = k.transpose(1, 2)
        vh = v.transpose(1, 2)
        out = _F.scaled_dot_product_attention(qh, kh, vh, attn_mask=attn_mask, dropout_p=p, scale=scale)
        return out.transpose(1, 2)  # [B, M, H, K]

    _xops.memory_efficient_attention = _mea_sdpa
    print("[blackwell] xformers.memory_efficient_attention rerouted to torch SDPA")
except Exception as _e:
    print(f"[blackwell] xformers SDPA shim skipped: {_e}")

import gradio as gr
from gradio_litmodel3d import LitModel3D

import shutil
from typing import *
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')
# TMP_DIR = "tmp/Trellis-demo"
# os.environ['GRADIO_TEMP_DIR'] = '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)

@spaces.GPU
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

@spaces.GPU
def preprocess_videos(video: str) -> List[Tuple[Image.Image, str]]:
    """
    Preprocess the input video for multi-image 3D generation.
    
    This function is called when a user uploads a video.
    It extracts frames from the video and processes each frame to prepare them
    for the multi-image 3D generation pipeline.
    
    Args:
        video (str): The path to the input video file
        
    Returns:
        List[Tuple[Image.Image, str]]: The list of preprocessed images ready for 3D generation
    """
    vid = imageio.get_reader(video, 'ffmpeg')
    fps = vid.get_meta_data()['fps']
    images = []
    for i, frame in enumerate(vid):
        if i % max(int(fps * 1), 1) == 0:
            img = Image.fromarray(frame)
            W, H = img.size
            img = img.resize((int(W / H * 512), 512))
            images.append(img)
    vid.close()
    processed_images = [pipeline.preprocess_image(image) for image in images]
    return processed_images

@spaces.GPU
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(duration=120)
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
    # import uuid
    # output_id = str(uuid.uuid4())
    # os.makedirs(f"{TMP_DIR}/{output_id}", exist_ok=True)
    # video_path = f"{TMP_DIR}/{output_id}/preview.mp4"
    # glb_path = f"{TMP_DIR}/{output_id}/mesh.glb"

    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]

# Create interface
demo = gr.Blocks(
    title="ReconViaGen",
    css="""
        .slider .inner { width: 5px; background: #FFF; }
        .viewport { aspect-ratio: 4/3; }
        .tabs button.selected { font-size: 20px !important; color: crimson !important; }
        h1, h2, h3 { text-align: center; display: block; }
        .md_feedback li { margin-bottom: 0px !important; }
    """
)
with demo:
    gr.Markdown("""
    # 💻 ReconViaGen
    <p align="center">
    <a title="Github" href="https://github.com/GAP-LAB-CUHK-SZ/ReconViaGen" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
        <img src="https://img.shields.io/github/stars/GAP-LAB-CUHK-SZ/ReconViaGen?label=GitHub%20%E2%98%85&logo=github&color=C8C" alt="badge-github-stars">
    </a>
    <a title="Website" href="https://jiahao620.github.io/reconviagen/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
        <img src="https://www.obukhov.ai/img/badges/badge-website.svg">
    </a>
    <a title="arXiv" href="https://jiahao620.github.io/reconviagen/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
        <img src="https://www.obukhov.ai/img/badges/badge-pdf.svg">
    </a>
    </p>
    
    ✨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="Input Video or Images", id=0) as multiimage_input_tab:
                    input_video = gr.Video(label="Upload Video", interactive=True, height=300)
                    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)

    input_video.upload(
        preprocess_videos,
        inputs=[input_video],
        outputs=[multiimage_prompt],
    )
    input_video.clear(
        lambda: tuple([None, None]),
        outputs=[input_video, multiimage_prompt],
    )
    multiimage_prompt.upload(
        preprocess_images,
        inputs=[multiimage_prompt],
        outputs=[multiimage_prompt],
    )

    generate_btn.click(
        get_seed,
        inputs=[randomize_seed, seed],
        outputs=[seed],
    ).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-v0-2")
    pipeline.cuda()
    pipeline.VGGT_model.cuda()
    pipeline.birefnet_model.cuda()
    demo.launch()