14 curated outputs from the release checkpoints. Rotate each object, separate its parts, or select a component to inspect it.
Generated part meshes with downstream material styling and hand-authored animation.
Two O-Voxel volumes preserve contact surfaces between parts. Connected components of each decoded volume form the part meshes.
On the 986 held-out objects completed by all six evaluated methods, KaiNinja achieves the best results on all seven reported whole-object (W) and Hungarian-matched part (P) metrics.
| Method | CDW↓ | F1W0.1 | F1W0.05 | mIoUP | CDP↓ | F1P0.1 | F1P0.05 |
|---|---|---|---|---|---|---|---|
| TRELLIS.2@1024 + X-Part | 0.0350 | 0.900 | 0.780 | 0.449 | 0.0804 | 0.703 | 0.559 |
| Hunyuan3D-2.1 + X-Part | 0.0314 | 0.917 | 0.814 | 0.489 | 0.0764 | 0.728 | 0.597 |
| OmniPart | 0.0366 | 0.901 | 0.769 | 0.456 | 0.0913 | 0.664 | 0.518 |
| PartPacker | 0.0389 | 0.892 | 0.749 | 0.449 | 0.0986 | 0.640 | 0.492 |
| AutoPartGen | 0.0373 | 0.901 | 0.771 | 0.466 | 0.0919 | 0.678 | 0.535 |
| KaiNinja (ours) | 0.0186 | 0.975 | 0.919 | 0.521 | 0.0613 | 0.785 | 0.692 |
Whole-object (W) and Hungarian-matched part (P) metrics; F1 thresholds are 0.1 and 0.05. Predictions are aligned to ground truth using the best of 48 signed axis permutations.
| Whole-object generators (reference) | CDW↓ | F1W0.1 | F1W0.05 | Fail %↓ |
|---|---|---|---|---|
| TRELLIS | 0.0340 | 0.916 | 0.807 | 3.9 |
| TRELLIS.2@512 | 0.0363 | 0.897 | 0.781 | 5.3 |
| TRELLIS.2@1024 | 0.0364 | 0.898 | 0.781 | 5.4 |
| TRELLIS.2@512, fine-tuned on our corpus | 0.0300 | 0.940 | 0.830 | 1.8 |
| Hunyuan3D-2.1 | 0.0340 | 0.912 | 0.807 | 6.8 |
| KaiNinja (ours, part-aware) | 0.0186 | 0.975 | 0.919 | 0.7 |
On this evaluation, KaiNinja improves whole-object fidelity over the listed generators, including TRELLIS.2 fine-tuned on the same corpus. Fail denotes the fraction of objects with CDW > 0.1.
@misc{yu2026kaininja,
title = {KaiNinja: Extending Native 3D Generators to the Part Level},
author = {Yu, Ruihan and Fu, Lian and Niu, Muyao and Huang, Zheng-hui and Tsai, Yu-Ju and Kuno, Sho and Lan, Fengbo and Yu, Yonghao and Wu, Erwin and Yang, Ming-Hsuan and Zhang, Kaipeng and Wang, Zhixiang},
year = {2026},
eprint = {2609.15659},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}
1 Alaya Lab2 The University of Tokyo3 University of California, Merced4 Institute of Science Tokyo