# JoVox

JoVox models all levels of a voxel octree jointly in one discrete flow-matching framework, so fine cells already generated take part in predicting coarse cells that are still masked, instead of coarse levels being fixed first and never revisited as in hierarchical methods. It also predicts cells two levels finer at once, halving the number of levels the network processes. On image-conditioned generation on Toys4k its voxel Chamfer distance is 0.01107 and IoU 0.4158, against 0.01250 and 0.4017 for TRELLIS.2, and with the same backbone it needs less GPU memory and has higher training throughput.

JoVox models all levels of a voxel octree jointly in one discrete flow-matching framework, so fine cells already generated take part in predicting coarse cells that are still masked, instead of coarse levels being fixed first and never revisited as in hierarchical methods. It also predicts cells two levels finer at once, halving the number of levels the network processes. On image-conditioned generation on Toys4k its voxel Chamfer distance is 0.01107 and IoU 0.4158, against 0.01250 and 0.4017 for TRELLIS.2, and with the same backbone it needs less GPU memory and has higher training throughput.