🚀 Fast-TRELLIS: Fast-SAM3D on TRELLIS [Official code repository]

May 22, 2026 · View on GitHub

TRELLIS Fast-SAM3D 3D Generation

Fast-TRELLIS is a TRELLIS implementation of the Fast-SAM3D acceleration framework.

Fast-TRELLIS Teaser

Fast-TRELLIS transfers the inference-time acceleration design of Fast-SAM3D to TRELLIS for efficient structured 3D generation.

📌 About This Repository

Fast-TRELLIS is based on Fast-SAM3D and ports its inference-time acceleration framework to TRELLIS. The migration covers the complete Fast-SAM3D design:

  • Modality-Aware Step Caching
  • Joint Spatiotemporal Token Carving
  • Spectral-Aware Token Aggregation

⭐ Star this repository if Fast-TRELLIS is helpful to your research or projects!

✨ Highlights

  • 🚀 TRELLIS acceleration: Built on top of microsoft/TRELLIS with Fast-SAM3D-style inference-time acceleration.
  • 🧩 Plug-and-play design: Supports original TRELLIS inference, TaylorSeer-style acceleration, and the full Fast-TRELLIS pipeline through simple runtime flags.
  • ⚡ Training-free speedup: Reduces TRELLIS inference latency while preserving the main geometry metrics.
  • 🖼️ Single-view and multi-view inference: Provides separate entry points for single-view generation and multi-view generation.
  • 🛠️ Official TRELLIS setup compatible: Environment setup, configuration files, and model checkpoint download follow the official TRELLIS repository.

📂 Settings

This repository focuses on fast TRELLIS inference. For the base model architecture, configuration format, environment setup, and checkpoint preparation, please follow the official microsoft/TRELLIS instructions.

Environment

Fast-TRELLIS uses the official TRELLIS environment. Please install dependencies according to:

👉 microsoft/TRELLIS - Installation

Models

Model download and checkpoint placement are fully aligned with the official TRELLIS release.Please follow:

👉 microsoft/TRELLIS - Model Download

After downloading, place the weights into the \checkpoints directory.

🚀 Quick Start

😍Gradio demo system

Fast-TRELLIS Teaser 

python app.py

🖼️ Single-View Inference

Use example for single-view inference:

# Original TRELLIS
python example.py

# TaylorSeer-style acceleration
python example.py --enable taylor

# Fast-TRELLIS acceleration
python example.py --enable faster --enable mesh

🧭 Multi-View Inference

Use example_multi for multi-view inference:

# Original TRELLIS multi-view inference
python example_multi_image.sh

# TaylorSeer-style acceleration
python example_multi_image.sh --enable taylor

# Fast-TRELLIS multi-view acceleration
python example_multi_image.sh --enable faster --enable mesh

📊 Experimental Results

Our method is motivated by inference-time heterogeneity in structured 3D generation, rather than by SAM3D-specific training objectives. Therefore, the key components are in principle portable to other voxel-/latent-based pipelines.

To directly evaluate this transferability, we migrated the full Fast-SAM3D framework to TRELLIS and evaluated it on Toys4K. This transfer covers the complete inference-time design of Fast-SAM3D, including Modality-Aware Step Caching, Joint Spatiotemporal Token Carving, and Spectral-Aware Token Aggregation.

MethodCD ↓F1@0.05 ↑vIoU ↑Latency (s) ↓GPU Memory (GB) ↓
TRELLIS0.063557.190.2957.68 (1.00×)10.38
+TaylorSeer0.063857.010.2994.65 (1.65×)10.40
+Fast3DCache0.065855.690.2487.91 (0.97×)10.52
+Ours0.063757.150.3003.40 (2.26×)9.97

On TRELLIS, our method reduces inference time from 7.68s to 3.40s (2.26×) while keeping the main geometry metrics essentially unchanged. It is also faster than TaylorSeer on this setup (4.65s) and avoids the stronger quality/runtime trade-off seen in Fast3DCache (CD 0.0658, F1 55.69, vIoU 0.248, 7.91s).

These results support that the Fast-SAM3D acceleration framework is not tied to SAM3D only, but is also applicable to related structured, iterative voxel-/latent-based 3D generation backbones.

🧠 Reference

Fast-SAM3D

@misc{feng2026fastsam3d3dfyimagesfaster,
      title={Fast-SAM3D: 3Dfy Anything in Images but Faster}, 
      author={Weilun Feng and Mingqiang Wu and Zhiliang Chen and Chuanguang Yang and Haotong Qin and Yuqi Li and Xiaokun Liu and Guoxin Fan and Zhulin An and Libo Huang and Yulun Zhang and Michele Magno and Yongjun Xu},
      year={2026},
      eprint={2602.05293},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.05293}, 
}

🙏 Acknowledgements

This project is based on the excellent TRELLIS repository and the Fast-SAM3D acceleration framework. We sincerely thank the authors and contributors for their inspiring work.

📄 License

This project is released under the MIT License.

📧 Contact

For questions or suggestions, please open an issue or contact:

State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences