✦
Summary
Block3D generates 3D shape geometry from text using block-wise diffusion. It combines the efficiency of processing only part of a shape representation at a time with the ability to revise uncertain predictions, addressing the latency and error-commitment tradeoffs of conventional autoregressive and full-sequence diffusion methods.
Key Features
Text prompts describe the desired 3D object.
The method uses a discrete sequence of 1,024 shape tokens.
Blocks are generated sequentially, with parallel denoising inside each active block.
Confidence-guided correction revises uncertain tokens before a block is finalized.
A frozen shape decoder converts the completed token sequence into geometry.
The evaluated setup averages 4.99 seconds end-to-end generation.
The reported speedup is 5.15 times against a fine-tuned autoregressive baseline.
The repository announces training, inference, and evaluation code.
About Block3D
The model partitions a 1,024-token shape sequence into contiguous blocks. Blocks are generated in causal order, while tokens within the active block are jointly denoised. Confidence-guided correction revisits uncertain tokens before a frozen shape decoder converts the completed sequence into geometry.
Block3D is useful for text-to-3D research and rapid asset prototyping. On the reported held-out TRELLIS-500K evaluation, it reduces mean end-to-end time from 25.71 to 4.99 seconds relative to the stated autoregressive baseline. Public code and model links support experimentation under the applicable upstream asset and model terms.