Step1X-3D Releases High-Quality 3D Asset Generation and Open Source Plan
On May 13, 2025, the Step1X-3D team officially announced the release of a high-quality 3D asset generation solution, along with open access to related models, datasets, and training/inference code. This project aims to provide easier and higher-quality 3D asset generation tools for creators and developers.
Key highlights of Step1X-3D include:
- Supports the generation of high-fidelity geometry and diverse texture maps, ensuring excellent alignment between geometry and textures.
- Provides a complete open-source technical report, inference code, training code, and model weights for free use and secondary development.
- Released a dataset of 800,000 high-quality 3D asset unique identifiers (uids) for training and synthesis.
- Offers an online demo (on Huggingface) for users to experience 3D asset generation directly.
- Future updates will support more controllable generation methods, such as multi-view, bounding box, and skeleton-based conditions.
Main Resources
- Online Demo (Huggingface web live)
- High-Quality 3D Asset Dataset
- Technical Report (arXiv)
- Project GitHub
Project Overview
Step1X-3D is dedicated to addressing challenges in 3D generation, such as data scarcity and algorithmic limitations. The team has built a dataset of 2 million high-quality 3D assets through a rigorous selection process and adopted a two-stage architecture: first generating 3D geometry, then synthesizing textures. This approach not only improves generation quality but also supports the transfer of 2D control techniques (like LoRA) to 3D generation.
All models, training, and inference code are open to the community for learning and use. Users without a technical background can also experience automatic 3D asset generation through the online demo or by downloading the models.