KaiNinja
Extending Native 3D Generators
to the Part Level

Paper arXiv Code Video BibTeX

KaiNinja extends TRELLIS.2 to generate separate part meshes from one image, using two O-Voxel volumes with no input masks or segmentation network.

Kaininja concept illustration: an armchair, speaker and reed diffuser reflected as pastel-colored parts in an illuminated mirror
INTERACTIVE RESULTS

Part-structured 3D generation

14 curated outputs from the release checkpoints. Rotate each object, separate its parts, or select a component to inspect it.

GENERATED ASSET

Dining chair

Input image
INPUT IMAGE
— parts
Loading…
Drag to rotate · scroll to zoom · click a part to inspect
Explode
APPLICATIONS

Part assembly, materials and motion

KAININJA IN MOTION · 55 SECONDS

Generated part meshes with downstream material styling and hand-authored animation.

Method

Complementary volumes, distinct parts

Two O-Voxel volumes preserve contact surfaces between parts. Connected components of each decoded volume form the part meshes.

— parts
Loading…
Drag to rotate · switch between volumes and individual parts
Explode
Results

Part-level and whole-object fidelity

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.

MethodCDWF1W0.1F1W0.05mIoUPCDPF1P0.1F1P0.05
TRELLIS.2@1024 + X-Part0.03500.9000.7800.4490.08040.7030.559
Hunyuan3D-2.1 + X-Part0.03140.9170.8140.4890.07640.7280.597
OmniPart0.03660.9010.7690.4560.09130.6640.518
PartPacker0.03890.8920.7490.4490.09860.6400.492
AutoPartGen0.03730.9010.7710.4660.09190.6780.535
KaiNinja (ours)0.01860.9750.9190.5210.06130.7850.692

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)CDWF1W0.1F1W0.05Fail %↓
TRELLIS0.03400.9160.8073.9
TRELLIS.2@5120.03630.8970.7815.3
TRELLIS.2@10240.03640.8980.7815.4
TRELLIS.2@512, fine-tuned on our corpus0.03000.9400.8301.8
Hunyuan3D-2.10.03400.9120.8076.8
KaiNinja (ours, part-aware)0.01860.9750.9190.7

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.

BibTeX

@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}
}

Team

1 Alaya Lab2 The University of Tokyo3 University of California, Merced4 Institute of Science Tokyo