Utonia — MLX (safetensors) weights for Apple Silicon
MLX-converted weights for Utonia (Utonia: Toward One Encoder for All Point Clouds, arXiv: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, a numerically verified Swift/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 (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
SubMConv3dkernels 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)
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:
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 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
@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}
}