JoVox-64

Image-conditioned 64³ voxel structure generator of JoVox: Joint Hierarchical Flow Matching for Lightweight Voxel Generation. JoVox is a discrete flow matching framework that learns the joint hierarchical distribution of voxels while keeping generation lightweight. One model generates all levels, 4³ → 16³ → 64³, from a single image.

Paper: arXiv:2611.10690 · Code: github.com/yuyujunjun/JoVox · Project page: yuyujunjun.github.io/Jovox-page

Usage

Install the code as described in its README, then download the weights:

hf download xiangjuntang/JoVox-64 model.safetensors --local-dir checkpoints/voxelgen64

Generate a structure and a mesh from your own images:

python scripts/img2mesh.py --stage prep  --in_dir <images> --out <out>
python scripts/img2mesh.py --stage voxel --out <out> --seed 0
python scripts/img2mesh.py --stage mesh  --out <out> --seed 0

Robustness to input images

Our training images were rendered from random viewpoints, but the size of the object in these images varies only within a limited range, and no crop augmentation was applied to widen it. As a result, the current model may not achieve optimal results when the object's size in the input image falls outside this range. Our image-to-mesh pipeline already mitigates this by centering the object in each input image and resizing it into this range before generation. We will open-source a model that is more robust to input images.

Model

File	Content
model.safetensors	Weights of the generator (EMA), used with configs/voxelgen64.yaml

The condition image is encoded by DINOv3, which is downloaded separately.

Training Data

The sketchfab, 3D-FUTURE, HSSD and ABO subsets of TRELLIS-500K. Toys4k was held out for testing.

Results

On 1000 Toys4k objects (image-conditioned):

Method	CD_V ↓	IoU ↑	CD_shell ↓	FID_N ↓	Uni3D ↑
TRELLIS.2	0.01250	0.4017	0.524e-3	33.3	0.3865
JoVox	0.01107	0.4158	0.407e-3	31.6	0.3868

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}
}