Researchers introduce Block3D, a text-to-3D generation framework that partitions discrete shape-token sequences into blocks, generates them sequentially, and jointly denoises all tokens within each block, incorporating confidence-guided intra-block correction that revises low-confidence tokens before a block is finalized. Evaluated on TRELLIS-500K, Block3D achieves a 5.15x speedup over autoregressive baselines, reducing generation time from 25.71 to 4.99 seconds while preserving geometric quality.
