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.
