Block3D: Efficient Text-to-3D Generation via Block-Wise Diffusion
Splits discrete shape-token sequences into contiguous blocks; generates blocks autoregressively while jointly denoising all tokens within each block. Confidence-guided intra-block correction revises low-confidence tokens before a block is finalized. On a held-out TRELLIS-500K subset, mean end-to-end generation time drops from 25.71 s to 4.99 s — a 5.15x speedup over a fine-tuned autoregressive baseline — while preserving geometric quality.
Why it matters
HF Daily 27 upvotes on Aug 25. Hybrid autoregressive + intra-block diffusion is a practical compromise for 3D generation that beats pure-AR baselines on speed without losing the global coherence of pure diffusion.
Importance: 2/5
HF Daily 27 upvotes
Sources
official
arXiv abstract
official
Project page
media
HF Daily Papers page