Block3D: Efficient Text-to-3D Generation via Block-Wise Diffusion
Block3D reports a 5.15x generation speedup while preserving geometric fidelity.
The paper proposes splitting a discrete 3D shape-token sequence into contiguous blocks, then generating those blocks autoregressively while denoising each block’s tokens together. It adds confidence-guided correction to revise low-confidence tokens before a block is locked in. On a held-out TRELLIS-500K set, the authors say mean end-to-end generation time fell from 25.71 seconds to 4.99 seconds versus a fine-tuned autoregressive baseline. HF Daily Papers' note
The paper proposes splitting a discrete 3D shape-token sequence into contiguous blocks, then generating those blocks autoregressively while denoising each block’s tokens together. It adds confidence-guided correction to revise low-confidence tokens before a block is locked in. On a held-out TRELLIS-500K set, the authors say mean end-to-end generation time fell from 25.71 seconds to 4.99 seconds versus a fine-tuned autoregressive baseline. HF Daily Papers' note
score 5