JoVox jointly models voxel hierarchy with discrete flow matching
Original title: JoVox: Joint Hierarchical Flow Matching for Lightweight Voxel Generation
Authors report 8x training throughput, 1/15 forward FLOPs, and 11% lower Chamfer distance than TRELLIS.2.
JoVox is a discrete flow-matching method that learns the joint distribution of every level in a voxel hierarchy so finer cells can inform coarse cells. A two-level predictor emits cells two levels finer at once, and an unmasking step revises early decisions after finer cells appear. The authors say it beats baselines on unconditional and image-conditioned voxel generation and report higher quality than TRELLIS.2 at lower compute.
Unconfirmed: Throughput, FLOPs, GPU counts, and Chamfer distance are the authors' comparisons with TRELLIS.2.
- Two-level prediction halves the number of hierarchy levels processed.
- Training throughput is reported as 8x that of TRELLIS.2.
- A forward pass uses 1/15 of TRELLIS.2 FLOPs.
- Training used 4 GPUs versus 32 for TRELLIS.2.
- Chamfer distance is reported 11% lower than TRELLIS.2.
From the original article
1 min read at huggingface.co
Generating high-fidelity voxels while keeping their lightweight nature remains an open problem in computer vision. Current generators are either high-fidelity but computationally heavy, or lightweight but of lower fidelity. Hierarchical methods keep the lightweight nature of voxels.
By Xiangjun Tang, Jianqi Chen, Peter Wonka · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.