app.py

# MIT License
# (see original notice and terms)

import gradio as gr
import os

# ---- Force CPU-only environment globally ----
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"          # hide GPUs from torch
os.environ.setdefault("ATTN_BACKEND", "sdpa")      # avoid xformers path
os.environ.setdefault("SPCONV_ALGO", "native")     # safe sparseconv algo
# ---------------------------------------------

from typing import *
import torch
import numpy as np
import tempfile
import zipfile
import types
import importlib

# ---------------------------------------------------------------------------
# Ensure bundled hi3dgen sources are available (extracted from hi3dgen.zip)
# ---------------------------------------------------------------------------
def _ensure_hi3dgen_available():
    pkg_name = 'hi3dgen'
    here = os.path.dirname(__file__)
    pkg_dir = os.path.join(here, pkg_name)
    if os.path.isdir(pkg_dir):
        return
    archive_path = os.path.join(here, f"{pkg_name}.zip")
    if not os.path.isfile(archive_path):
        raise FileNotFoundError(
            f"Required archive {archive_path} is missing. Upload hi3dgen.zip next to app.py."
        )
    try:
        with zipfile.ZipFile(archive_path, 'r') as zf:
            zf.extractall(here)
    except Exception as e:
        raise RuntimeError(f"Failed to extract {archive_path}: {e}")

_ensure_hi3dgen_available()

# ---------------------------------------------------------------------------
# xformers stub (CPU-friendly fallback for xformers.ops.memory_efficient_attention)
# ---------------------------------------------------------------------------
def _ensure_xformers_stub():
    import sys
    if 'xformers.ops' in sys.modules:
        return
    import torch.nn.functional as F
    xf_mod = types.ModuleType('xformers')
    ops_mod = types.ModuleType('xformers.ops')

    def memory_efficient_attention(query, key, value, attn_bias=None):
        return F.scaled_dot_product_attention(query, key, value, attn_bias)

    ops_mod.memory_efficient_attention = memory_efficient_attention
    xf_mod.ops = ops_mod
    sys.modules['xformers'] = xf_mod
    sys.modules['xformers.ops'] = ops_mod

_ensure_xformers_stub()

# ---------------------------------------------------------------------------
# Import pipeline AFTER stubbing xformers, then patch CUDA-hotspots to CPU
# ---------------------------------------------------------------------------
from hi3dgen.pipelines import Hi3DGenPipeline
import trimesh

# ---- Force CPU inside hi3dgen (avoid any CUDA paths) ----
print("[PATCH] Applying CPU monkey-patches to hi3dgen")

# 1) utils_cube.construct_dense_grid(..., device=...) -> force CPU
uc = importlib.import_module("hi3dgen.representations.mesh.utils_cube")
if not hasattr(uc, "_CPU_PATCHED"):
    _orig_construct_dense_grid = uc.construct_dense_grid
    def _construct_dense_grid_cpu(res, device=None):
        return _orig_construct_dense_grid(res, device="cpu")
    uc.construct_dense_grid = _construct_dense_grid_cpu
    uc._CPU_PATCHED = True
    print("[PATCH] utils_cube.construct_dense_grid -> CPU")

# 2) cube2mesh.EnhancedMarchingCubes default device -> force CPU (flexible)
cm = importlib.import_module("hi3dgen.representations.mesh.cube2mesh")
M = cm.EnhancedMarchingCubes
if not hasattr(M, "_CPU_PATCHED"):
    _orig_init = M.__init__
    def _init_cpu(self, *args, **kwargs):
        # Ensure device is CPU regardless of how it's passed
        if "device" in kwargs:
            kwargs["device"] = torch.device("cpu")
        else:
            kwargs.setdefault("device", torch.device("cpu"))
        return _orig_init(self, *args, **kwargs)
    M.__init__ = _init_cpu
    M._CPU_PATCHED = True
    print("[PATCH] cube2mesh.EnhancedMarchingCubes.__init__ -> CPU (flex)")

# 3) Belt & suspenders: coerce torch.arange(device='cuda') to CPU if any call slips through
if not hasattr(torch, "_ARANGE_CPU_PATCHED"):
    _orig_arange = torch.arange
    def _arange_cpu(*args, **kwargs):
        dev = kwargs.get("device", None)
        if dev is not None and str(dev).startswith("cuda"):
            kwargs["device"] = "cpu"
        return _orig_arange(*args, **kwargs)
    torch.arange = _arange_cpu
    torch._ARANGE_CPU_PATCHED = True
    print("[PATCH] torch.arange(device='cuda') -> CPU")
# ----------------------------------------------------------

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

# ---------------------------------------------------------------------------
# Weights caching
# ---------------------------------------------------------------------------
def cache_weights(weights_dir: str) -> dict:
    from huggingface_hub import snapshot_download
    os.makedirs(weights_dir, exist_ok=True)
    model_ids = [
        "Stable-X/trellis-normal-v0-1",
        "Stable-X/yoso-normal-v1-8-1",
        "ZhengPeng7/BiRefNet",
    ]
    cached_paths = {}
    for model_id in model_ids:
        print(f"Caching weights for: {model_id}")
        local_path = os.path.join(weights_dir, model_id.split("/")[-1])
        if os.path.exists(local_path):
            print(f"Already cached at: {local_path}")
            cached_paths[model_id] = local_path
            continue
        print(f"Downloading and caching model: {model_id}")
        local_path = snapshot_download(
            repo_id=model_id,
            local_dir=os.path.join(weights_dir, model_id.split("/")[-1]),
            force_download=False
        )
        cached_paths[model_id] = local_path
        print(f"Cached at: {local_path}")
    return cached_paths

# ---------------------------------------------------------------------------
# Pre/Post processing and generation
# ---------------------------------------------------------------------------
def preprocess_mesh(mesh_prompt):
    print("Processing mesh")
    trimesh_mesh = trimesh.load_mesh(mesh_prompt)
    out_path = mesh_prompt + '.glb'
    trimesh_mesh.export(out_path)
    return out_path

def preprocess_image(image):
    if image is None:
        return None
    return hi3dgen_pipeline.preprocess_image(image, resolution=1024)

def generate_3d(
    image,
    seed: int = -1,
    ss_guidance_strength: float = 3,
    ss_sampling_steps: int = 50,
    slat_guidance_strength: float = 3,
    slat_sampling_steps: int = 6,
):
    if image is None:
        return None, None, None

    if seed == -1:
        seed = np.random.randint(0, MAX_SEED)

    image = hi3dgen_pipeline.preprocess_image(image, resolution=1024)
    normal_image = normal_predictor(
        image,
        resolution=768,
        match_input_resolution=True,
        data_type='object'
    )

    outputs = hi3dgen_pipeline.run(
        normal_image,
        seed=seed,
        formats=["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,
        },
    )
    generated_mesh = outputs['mesh'][0]

    import datetime
    output_id = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
    os.makedirs(os.path.join(TMP_DIR, output_id), exist_ok=True)
    mesh_path = f"{TMP_DIR}/{output_id}/mesh.glb"

    trimesh_mesh = generated_mesh.to_trimesh(transform_pose=True)
    trimesh_mesh.export(mesh_path)

    return normal_image, mesh_path, mesh_path

def convert_mesh(mesh_path, export_format):
    if not mesh_path:
        return None
    temp_file = tempfile.NamedTemporaryFile(suffix=f".{export_format}", delete=False)
    temp_file_path = temp_file.name
    mesh = trimesh.load_mesh(mesh_path)
    mesh.export(temp_file_path)
    return temp_file_path

# ---------------------------------------------------------------------------
# UI
# ---------------------------------------------------------------------------
with gr.Blocks(css="footer {visibility: hidden}") as demo:
    gr.Markdown(
        """
        <h1 style='text-align: center;'>Hi3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging</h1>
        <p style='text-align: center;'>
            <strong>V0.1, Introduced By 
            <a href="https://gaplab.cuhk.edu.cn/" target="_blank">GAP Lab</a> (CUHKSZ) and 
            <a href="https://www.nvsgames.cn/" target="_blank">Game-AIGC Team</a> (ByteDance)</strong>
        </p>
        """
    )

    with gr.Row():
        gr.Markdown("""
            <p align="center">
            <a title="Website" href="https://stable-x.github.io/Hi3DGen/" 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://stable-x.github.io/Hi3DGen/hi3dgen_paper.pdf" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
                <img src="https://www.obukhov.ai/img/badges/badge-pdf.svg">
            </a>
            <a title="Github" href="https://github.com/Stable-X/Hi3DGen" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
                <img src="https://img.shields.io/github/stars/Stable-X/Hi3DGen?label=GitHub%20%E2%98%85&logo=github&color=C8C" alt="badge-github-stars">
            </a>
            <a title="Social" href="https://x.com/ychngji6" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
                <img src="https://www.obukhov.ai/img/badges/badge-social.svg" alt="social">
            </a>
            </p>
        """)

    with gr.Row():
        with gr.Column(scale=1):
            with gr.Tabs():
                with gr.Tab("Single Image"):
                    with gr.Row():
                        image_prompt = gr.Image(label="Image Prompt", image_mode="RGBA", type="pil")
                        normal_output = gr.Image(label="Normal Bridge", image_mode="RGBA", type="pil")
                with gr.Tab("Multiple Images"):
                    gr.Markdown(
                        "<div style='text-align: center; padding: 40px; font-size: 24px;'>Multiple Images functionality is coming soon!</div>"
                    )

            with gr.Accordion("Advanced Settings", open=False):
                seed = gr.Slider(-1, MAX_SEED, label="Seed", value=0, step=1)
                gr.Markdown("#### Stage 1: Sparse Structure Generation")
                with gr.Row():
                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3, step=0.1)
                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=50, 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=6, step=1)

            with gr.Group():
                with gr.Row():
                    gen_shape_btn = gr.Button("Generate Shape", size="lg", variant="primary")

        with gr.Column(scale=1):
            with gr.Column():
                model_output = gr.Model3D(label="3D Model Preview (Each model is ~40MB; may take ~1 min to load)")
            with gr.Column():
                export_format = gr.Dropdown(
                    choices=["obj", "glb", "ply", "stl"],
                    value="glb",
                    label="File Format"
                )
                download_btn = gr.DownloadButton(label="Export Mesh", interactive=False)

    image_prompt.upload(
        preprocess_image,
        inputs=[image_prompt],
        outputs=[image_prompt]
    )

    gen_shape_btn.click(
        generate_3d,
        inputs=[
            image_prompt, seed,
            ss_guidance_strength, ss_sampling_steps,
            slat_guidance_strength, slat_sampling_steps
        ],
        outputs=[normal_output, model_output, download_btn]
    ).then(
        lambda: gr.Button(interactive=True),
        outputs=[download_btn],
    )

    def update_download_button(mesh_path, export_format):
        if not mesh_path:
            return gr.File.update(value=None, interactive=False)
        download_path = convert_mesh(mesh_path, export_format)
        return download_path

    export_format.change(
        update_download_button,
        inputs=[model_output, export_format],
        outputs=[download_btn]
    ).then(
        lambda: gr.Button(interactive=True),
        outputs=[download_btn],
    )

    examples = None

    gr.Markdown(
        """
        **Acknowledgments**: Hi3DGen is built on the shoulders of giants. We would like to express our gratitude to the open-source research community and the developers of these pioneering projects:
        - **3D Modeling:** Our 3D Model is finetuned from the SOTA open-source 3D foundation model [Trellis](https://github.com/microsoft/TRELLIS); inspired by Rodin, Tripo, and Dora.
        - **Normal Estimation:** Builds on StableNormal and GenPercept.
        """
    )

# ---------------------------------------------------------------------------
# Entry
# ---------------------------------------------------------------------------
if __name__ == "__main__":
    # Cache model weights locally
    cache_weights(WEIGHTS_DIR)

    # Load pipeline on CPU
    hi3dgen_pipeline = Hi3DGenPipeline.from_pretrained("weights/trellis-normal-v0-1")
    try:
        hi3dgen_pipeline.to("cpu")
    except Exception:
        pass  # some pipelines may not implement .to

    # Initialize normal predictor (CPU)
    try:
        normal_predictor = torch.hub.load(
            os.path.join(torch.hub.get_dir(), 'hugoycj_StableNormal_main'),
            "StableNormal_turbo",
            yoso_version='yoso-normal-v1-8-1',
            source='local',
            local_cache_dir='./weights',
            pretrained=True
        )
    except Exception:
        normal_predictor = torch.hub.load(
            "hugoycj/StableNormal",
            "StableNormal_turbo",
            trust_repo=True,
            yoso_version='yoso-normal-v1-8-1',
            local_cache_dir='./weights'
        )
    try:
        normal_predictor.to("cpu")
    except Exception:
        pass

    # Launch the Gradio app
    demo.launch(share=False, server_name="0.0.0.0")