README.md

---
license: cc-by-nc-4.0
library_name: mlx
pipeline_tag: graph-ml
tags:
- mlx
- mlx-swift
- 3d
- point-cloud
- self-supervised-learning
- point-transformer-v3
base_model: Pointcept/Utonia
---

# Utonia — MLX (safetensors) weights for Apple Silicon

MLX-converted weights for **[Utonia](https://huggingface.co/Pointcept/Utonia)** (*Utonia: Toward One Encoder for All Point Clouds*, [arXiv:2603.03283](https://arxiv.org/abs/2603.03283)) — a Point Transformer V3 (mode-3) encoder pretrained across indoor RGB-D, outdoor LiDAR, remote sensing, object CAD, and video-lifted point clouds.

These weights power **[mlx-swift-utonia](https://github.com/mnmly/mlx-swift-utonia)**, a numerically verified Swift/[mlx-swift](https://github.com/ml-explore/mlx-swift) port that runs Utonia natively on Apple Silicon (macOS): full-resolution multi-million-point clouds, ScanNet-20 semantic segmentation, and PCA feature visualization.

## Files

| File | Contents |
| --- | --- |
| `utonia.safetensors` | Encoder weights (137,253,816 params, float32) — converted from `utonia.pth` |
| `utonia_config.json` | Model config extracted from the checkpoint (`enc_channels=(54,108,216,432,576)`, `enc_depths=(3,3,3,12,3)`, `enc_num_head=(3,6,12,24,32)`, 3D RoPE, 4 serialization curves) |
| `utonia_seg_head_sc.safetensors` | ScanNet-20 linear-probe segmentation head (`Linear(1386→20)`) — converted from `utonia_linear_prob_head_sc.pth` |
| `utonia_seg_head_sc_config.json` | Seg-head config |
| `*_manifest.txt` | Tensor name/shape manifests for both checkpoints |

## Changes from the original

Converted from the PyTorch checkpoints in [Pointcept/Utonia](https://huggingface.co/Pointcept/Utonia) (no retraining, no fine-tuning — the numbers are byte-identical modulo the format changes below):

- `.pth` (pickled state dict) → `.safetensors`, float32.
- Config dict extracted from the checkpoint into a standalone JSON.
- Tensor **values are unchanged**; the Swift loader transposes the spconv `SubMConv3d` kernels from `(C_out, k, k, k, C_in)` to `(k, k, k, C_in, C_out)` at load time.

Numerical parity of the Swift port against the reference PyTorch implementation: bit-exact serialization, encoder relative error 0.27 % (fp32 GPU drift), 99.87 % semantic-segmentation argmax agreement.

## Usage (mlx-swift)

```swift
import Utonia

let session = try UtoniaSession.load(SessionConfig(weightsDir: weightsDirURL))
let result = session.run(RawCloud(coord: coords, color: colors, normal: normals))
```

Or via the CLI from [mlx-swift-utonia](https://github.com/mnmly/mlx-swift-utonia):

```sh
utonia-cli semseg --weights-dir weights --input scene/ --output segmented.ply
utonia-cli pca    --weights-dir weights --input scene/ --output features.ply
```

## License

**CC-BY-NC-4.0**, inherited from the original [Pointcept/Utonia](https://huggingface.co/Pointcept/Utonia) weights — **non-commercial use only**. The mlx-swift-utonia *code* is licensed separately (see its repository); this restriction applies to the weights.

All credit for the model belongs to the Utonia authors (Pointcept / The University of Hong Kong and collaborators).

## Citation

```bibtex
@misc{zhang2026utoniaencoderpointclouds,
      title={Utonia: Toward One Encoder for All Point Clouds},
      author={Yujia Zhang and Xiaoyang Wu and Yunhan Yang and Xianzhe Fan and Han Li and Yuechen Zhang and Zehao Huang and Naiyan Wang and Hengshuang Zhao},
      year={2026},
      eprint={2603.03283},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.03283},
}

@misc{pointcept2023,
    title={Pointcept: A Codebase for Point Cloud Perception Research},
    author={Pointcept Contributors},
    howpublished = {\url{https://github.com/Pointcept/Pointcept}},
    year={2023}
}
```