TripoSF represents a significant leap forward in 3D shape modeling, combining high-resolution capabilities with arbitrary topology support. Our approach enables:

• 📈 Ultra-high resolution mesh modeling (up to ${1024}^3$)
• 🎯 Direct optimization from rendering losses
• 🌐 Efficient handling of open surfaces and complex topologies
• 💾 Dramatic memory reduction through sparse computation
• 🔄 Differentiable mesh extraction with sharp features

SparseFlex, the core design powering TripoSF, introduces a sparse voxel structure that:

• Focuses computational resources only on surface-adjacent regions
• Enables natural handling of open surfaces (like cloth or leaves)
• Supports complex internal structures without compromises
• Achieves massive memory reduction compared to dense representations

• [2025-03] Initial Release:
  • Pretrained VAE model weights (${1024}^3$ reconstruction)
  • Inference scripts and examples
  • SparseFlex implementation

• CUDA-capable GPU (≥12GB VRAM for ${1024}^3$ resolution)
• PyTorch 2.0+

1. Clone the repository:

git clone https://github.com/VAST-AI-Research/TripoSF.git
cd TripoSF

2. Install dependencies:

# Install PyTorch (select the correct CUDA version)
pip install torch torchvision --index-url https://download.pytorch.org/whl/{your-cuda-version}

# Install other dependencies
pip install -r requirements.txt

1. Download our pretrained models from Hugging Face
2. Place the models in the ckpts/ directory

Basic reconstruction using TripoSFVAE:

python inference.py --mesh-path "assets/examples/jacket.obj" \
                   --output-dir "outputs/" \
                   --config "configs/TripoSFVAE_1024.yaml"

python app.py

• Enable pruning in the configuration:

  pruning: true
• Benefits:
  • Higher-fidelity reconstruction
  • Faster processing
  • Better memory efficiency

• Increase sampling density:

  sample_points_num: 1638400  # Default: 819200
• Adjust resolution based on detail requirements:

  resolution: 1024  # Options: 256, 512, 1024

TripoSF VAE Architecture:

• Input: Point clouds (preserving source geometry details)
• Encoder: Sparse transformer for efficient geometry encoding
• Decoder: Self-pruning upsampling modules maintaining sparsity
• Output: High-resolution SparseFlex parameters for mesh extraction

@article{he2025triposf,
  title={SparseFlex: High-Resolution and Arbitrary-Topology 3D Shape Modeling},
  author={He, Xianglong and Zou, Zi-Xin and Chen, Chia-Hao and Guo, Yuan-Chen and Liang, Ding and Yuan, Chun and Ouyang, Wanli and Cao, Yan-Pei and Li, Yangguang},
  journal={arXiv preprint arXiv:2503.21732},
  year={2025}
}

Our work builds upon these excellent repositories:

• Trellis
• Flexicubes

This project is licensed under the MIT License - see the LICENSE file for details.