JoVox: Joint Hierarchical Flow Matching for Lightweight Voxel Generation
JoVox reports 8× higher training throughput than TRELLIS.2, with an 11% lower Chamfer distance.
The framework lets finer voxel cells inform coarse cells, addressing early placement errors that hierarchical generators typically leave uncorrected. It learns all hierarchy levels jointly and predicts two levels finer at once, halving the number of levels the network processes. An unmasking strategy revisits early decisions to correct inconsistencies as finer cells appear. The authors report better unconditional and image-conditioned generation, using four GPUs versus 32 and one-fifteenth of the forward-pass FLOPs of TRELLIS.2.
HF Daily Papers' note
The framework lets finer voxel cells inform coarse cells, addressing early placement errors that hierarchical generators typically leave uncorrected. It learns all hierarchy levels jointly and predicts two levels finer at once, halving the number of levels the network processes. An unmasking strategy revisits early decisions to correct inconsistencies as finer cells appear. The authors report better unconditional and image-conditioned generation, using four GPUs versus 32 and one-fifteenth of the forward-pass FLOPs of TRELLIS.2.
HF Daily Papers' note
score 5