TL;DR of the paper, generated automatically.

JoVox is a new voxel generator that predicts all hierarchy levels together, not one at a time.

• Speed: 8x training throughput and 1/15 the FLOPs per forward pass, vs. TRELLIS.2.
• Training: trained on 4 H100 GPUs instead of 32, per the authors.
• Quality: 11% lower Chamfer distance on 1000 Toys4k objects, image-conditioned.
• Release: code, 64³ model and image-to-mesh script are out.