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.

Demo: Hugging Face Space · 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 a model (see Models):

hf download xiangjuntang/JoVox-64 model.safetensors config.yaml --local-dir checkpoints/voxelgen64   # JoVox-64
hf download xiangjuntang/JoVox-64 --include "jovox64_robust/*" --local-dir checkpoints                  # JoVox-64-Robust

Generate a structure and a mesh from your own images, with <model> the folder you downloaded (checkpoints/voxelgen64 or checkpoints/jovox64_robust):

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

The config.yaml of each model sets how prep frames the input image for that model.

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, JoVox-64 may not achieve optimal results when the object's size in the input image falls outside this range. Our image-to-mesh pipeline mitigates this by centering the object in each input image and resizing it into this range before generation.

JoVox-64-Robust is an update of JoVox-64 for such images. It is less sensitive to the size of the object in the input image, and we recommend it for your own images.

Models

Model Files Content
JoVox-64 model.safetensors, config.yaml The model evaluated in the paper (weights of the generator, EMA)
JoVox-64-Robust jovox64_robust/model.safetensors, jovox64_robust/config.yaml The update for input images in which the object appears at different sizes

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}
}
Downloads last month
21
Safetensors
Model size
2B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Space using xiangjuntang/JoVox-64 1

Paper for xiangjuntang/JoVox-64