README.md

---
license: mit
tags:
- image-to-3d
- 3d-aigc
- 3d-reconstruction
- 3d-models
- 3d-generation
---


# Direct3D‑S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention

<div align="center">
  <a href=https://www.neural4d.com/research/direct3d-s2 target="_blank"><img src=https://img.shields.io/badge/Project%20Page-333399.svg?logo=googlehome height=22px></a>
  <a href=https://huggingface.co/spaces/wushuang98/Direct3D-S2-v1.0-demo target="_blank"><img src=https://img.shields.io/badge/%F0%9F%A4%97%20Demo-276cb4.svg height=22px></a>
  <a href=https://huggingface.co/spaces/wushuang98/Direct3D-S2-v1.0-demo target="_blank"><img src=https://img.shields.io/badge/%F0%9F%A4%97%20Models-d96902.svg height=22px></a>
  <a href=https://arxiv.org/pdf/2505.17412 target="_blank"><img src=https://img.shields.io/badge/Arxiv-b5212f.svg?logo=arxiv height=22px></a>
</div>

<div style="background: #fff; box-shadow: 0 4px 12px rgba(0,0,0,.15); display: inline-block; padding: 0px;">
    <img id="teaser" src="assets/teaserv6.png" alt="Teaser image of Direct3D-S2"/>
</div>

---

## ✨ News

- May 30, 2025: 🤯 We have released both v1.0 and v1.1. The new model offers even greater speed compared to FlashAttention-2, with **12.2×** faster forward pass and **19.7×** faster backward pass, resulting in nearly **2×** inference speedup over v1.0.
- May 30, 2025: 🔨 Release inference code and model.
- May 26, 2025: 🎁 Release live demo on 🤗 [Hugging Face](https://huggingface.co/spaces/wushuang98/Direct3D-S2-v1.0-demo).
- May 26, 2025: 🚀 Release paper and project page.

## 📝 Abstract

Generating high-resolution 3D shapes using volumetric representations such as Signed Distance Functions (SDFs) presents substantial computational and memory challenges. We introduce <strong class="has-text-weight-bold">Direct3D‑S2</strong>, a scalable 3D generation framework based on sparse volumes that achieves superior output quality with dramatically reduced training costs. Our key innovation is the <strong class="has-text-weight-bold">Spatial Sparse Attention (SSA)</strong> mechanism, which greatly enhances the efficiency of Diffusion Transformer (DiT) computations on sparse volumetric data. SSA allows the model to effectively process large token sets within sparse volumes, substantially reducing computational overhead and achieving a <em>3.9&times;</em> speedup in the forward pass and a <em>9.6&times;</em> speedup in the backward pass. Our framework also includes a variational autoencoder (VAE) that maintains a consistent sparse volumetric format across input, latent, and output stages. Compared to previous methods with heterogeneous representations in 3D VAE, this unified design significantly improves training efficiency and stability. Our model is trained on public available datasets, and experiments demonstrate that <strong class="has-text-weight-bold">Direct3D‑S2</strong> not only surpasses state-of-the-art methods in generation quality and efficiency, but also enables <strong class="has-text-weight-bold">training at 1024<sup>3</sup>  resolution with just 8 GPUs</strong>, a task typically requiring at least 32 GPUs for volumetric representations at 256<sup>3</sup> resolution, thus making gigascale 3D generation both practical and accessible.

## 🌟 Highlight

- **Gigascale 3D Generation**: Direct3D-S2 enables training at 1024<sup>3</sup> resolution with only 8 GPUs.
- **Spatial Sparse Attention (SSA)**: A novel attention mechanism designed for sparse volumetric data, enabling efficient processing of large token sets.
- **Unified Sparse VAE**: A variational autoencoder that maintains a consistent sparse volumetric format across input, latent, and output stages, improving training efficiency and stability.

## 🚀 Getting Started

### Installation

```sh
git clone https://github.com/DreamTechAI/Direct3D-S2.git

cd Direct3D-S2

pip install -r requirements.txt

pip install -e .

```

### Usage

```python
from direct3d_s2.pipeline import Direct3DS2Pipeline
pipeline = Direct3DS2Pipeline.from_pretrained(
  'wushuang98/Direct3D-S2', 
  subfolder="direct3d-s2-v-1-1"
).to("cuda:0")

mesh = pipeline(
  'assets/test/13.png', 
  sdf_resolution=1024, # 512 or 1024
  remesh=False, # Switch to True if you need to reduce the number of triangles.
)["mesh"]

mesh.export('output.obj')
```

### Web Demo

We provide a Gradio web demo for Direct3D-S2, which allows you to generate 3D meshes from images interactively.

```bash
python app.py
```

## 🤗 Acknowledgements

Thanks to the following repos for their great work, which helps us a lot in the development of Direct3D-S2:

- [Trellis](https://github.com/microsoft/TRELLIS)
- [SparseFlex](https://github.com/VAST-AI-Research/TripoSF)
- [native-sparse-attention-triton](https://github.com/XunhaoLai/native-sparse-attention-triton)
- [diffusers](https://github.com/huggingface/diffusers)

## 📄 License

Direct3D-S2 is released under the MIT License. See [LICENSE](LICENSE) for details.

## 📖 Citation

If you find our work useful, please consider citing our paper:

```bibtex
@article{wu2025direct3ds2gigascale3dgeneration,
  title={Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention}, 
  author={Shuang Wu and Youtian Lin and Feihu Zhang and Yifei Zeng and Yikang Yang and Yajie Bao and Jiachen Qian and Siyu Zhu and Philip Torr and Xun Cao and Yao Yao},
  journal={arXiv preprint arXiv:2505.17412},
  year={2025}
}
```