JoVox: Joint Hierarchical Flow Matching for Lightweight Voxel Generation

Xiangjun TangJianqi ChenPeter Wonka

King Abdullah University of Science and Technology (KAUST)

Paper (arXiv) Code Model

8×training throughputvs TRELLIS.2, same backbone
1/15FLOPs per forward pass0.82 vs 12.2 TFLOPs
4 vs. 32H100 GPUs to trainours: about 2.5 days to match TRELLIS.2
−11%Chamfer distancevs TRELLIS.2, image-conditioned

The joint distribution of all levels

HierarchicalLevel by level

p(O₀)

p(O₂ | O₁)

p(O_L | O_L−1)

level 0

level 2

level 4

OursAll levels jointly

p(O₀, …, O_L)

level 0

level 2

level 4

Efficiency

GPUs used for training

Per shape, one forward pass

One training step on one A100 (80 GB), TRELLIS.2 with the same backbone

OursTRELLIS.2× out of memory

Method walkthrough

Best viewed in full screen.

Results

Image-conditioned

Unconditional

Objects from the Toys4k test set. Drag any shape to rotate all of them.

Generated cells, distance to the ground truth (cells) < 22–33–44–66–8≥ 8

Missing parts, distance to the generated cells (cells) 2–33–44–66–8≥ 8

Image-conditioned generation on Toys4k (^∗ voxels from voxelized meshes; ^‡ evaluated on their own meshes). Best in bold, second best underlined.

Method	Voxels				Meshes
Method	CD_V ↓	IoU ↑	LPIPS_V ↓	DINO_V ↑	CD_shell (×10⁻³) ↓	LPIPS_N ↓	CLIP_N ↑	Uni3D ↑
TripoSR^∗‡	0.01663	0.3199	0.1837	0.7813	0.781	0.2291	0.7431	0.3414
TripoSG^∗‡	0.01325	0.3722	0.1626	0.8523	0.525	0.1698	0.7592	0.3849
Direct3D-S2^‡	0.01246	0.3694	0.1709	0.8550	0.539	0.1939	0.7600	0.3824
DVD	0.01973	0.2786	0.2119	0.7914	1.070	0.2412	0.7538	0.3718
TRELLIS	0.01328	0.3595	0.1641	0.8385	0.524	0.1899	0.7583	0.3799
TRELLIS.2	0.01250	0.4017	0.1524	0.8513	0.524	0.1666	0.7613	0.3865
Ours	0.01107	0.4158	0.1395	0.8727	0.407	0.1621	0.7615	0.3868

Our framework extended to 256³ (L = 6). This model is fine-tuned from L = 4 for only 32K steps and has not converged.

Pick an input image below. Drag either shape to rotate both.

Input

64³ level

256³

Citation

@article{tang2026jovox,
  title={JoVox: Joint Hierarchical Flow Matching for Lightweight Voxel Generation},
  author={Tang, Xiangjun and Chen, Jianqi and Wonka, Peter},
  journal={arXiv preprint arXiv:2611.10690},
  year={2026}
}