Spaces:
Running
Running
Submission: TripoSplat MLX for mlx-community
#38
by Jup33QE - opened
Submission: TripoSplat MLX for mlx-community
Project Overview
TripoSplat MLX is a native Apple Silicon (MLX) implementation of TripoSplat, a single-image-to-3D-Gaussian-Splat model by VAST-AI-Research.
This project ports the core TripoSplat submodules from PyTorch to MLX, enabling efficient on-device inference for 3D content creation on Mac.
What Has Been Done
MLX Ports (Complete)
| Component | Original | MLX Port | Status |
|---|---|---|---|
| DinoV3 ViT-H/16+ | PyTorch | mlx_models/dinov3_mlx.py |
✅ Working |
| Flux2 VAE Encoder | PyTorch | mlx_models/flux2vae_mlx.py |
✅ Working |
| LatentSeqMMFlowModel | PyTorch | mlx_models/flow_mlx.py |
✅ Working |
| OctreeGaussianDecoder | PyTorch | — | ⏳ Remains on MPS (complex dynamic sampling) |
Key Features
- Native MLX: No PyTorch dependency for encoders and flow model
- Weight-compatible: Loads original safetensors checkpoints directly
- Numerically verified: Outputs match PyTorch reference within tolerance
- End-to-end pipeline: Includes FLUX.2-klein-9B → BiRefNet → TripoSplat → SPZ export
- MIT License: Fully open source
Links
- HuggingFace Model: https://huggingface.co/Jup33QE/tripoflux-mlx
- GitHub Project: https://github.com/Jup33Q/tripoflux-mlx
- Original TripoSplat: https://github.com/VAST-AI-Research/TripoSplat
Request
We would like to contribute this MLX implementation to the mlx-community organization on HuggingFace. The model weights and MLX code are ready for transfer.
If the maintainers are interested, we can:
- Transfer the HuggingFace repository to
mlx-community/triposplat-mlx - Submit a PR to the mlx-explore/mlx-examples repository with the model implementations
- Provide any additional verification or benchmarks needed
Contact
- Author: Jup33Q (zxgzg@163.com)
- HuggingFace: https://huggingface.co/Jup33QE
- GitHub: https://github.com/Jup33Q
Submitted: 2026-07-23