Block3D is a text-to-3D generation method released on GitHub by ZIP Lab, accompanying an arXiv paper (2608.19567) and a project page. Instead of generating discrete shape tokens strictly one at a time, it reorganizes the token sequence into contiguous blocks: it generates one block at a time, denoises every token inside the active block in parallel, and permits bounded correction of already-placed tokens before a block is committed, aiming to keep autoregressive-style sample quality while cutting token-by-token decoding latency. It reuses a frozen Cube3D shape/text encoder for the condition sequence and a frozen Cube3D shape decoder to turn the finished token sequence into a mesh, rather than training these from scratch; the authors describe the work as built directly on Roblox's released Cube3D-v0.1 tokenizer, decoder, and GPT backbone. On 2026-08-24 the authors released training, inference, and evaluation code, not just the paper: the pipeline splits the TRELLIS-500K paired text-mesh dataset (not bundled) into a 300K-example training set and a fixed 100-object evaluation set, trains across four A100 80GB GPUs via torchrun for both the full model and a mask-to-token-only ablation, and ships an inference script plus an evaluation harness computing Chamfer distance, normal consistency, and F-score geometry metrics alongside an eight-view CLIPScore rendered with PyTorch3D. On that fixed evaluation set the authors report an internal ablation between the mask-to-token-only variant and the full mask-to-token-plus-correction variant (CD-L1 0.0813 vs 0.0775, F-score@1% 0.2872 vs 0.3088), not a comparison against an external baseline, and separately report 4.99 seconds mean end-to-end latency for the complete model on a single A100 80GB GPU. The most important caveat here is licensing: Block3D ships under a 'BLOCK3D DERIVATIVE WORK NOTICE' wrapping Roblox's Cube3D Research-Only RAIL-MS license, restricting use of code, weights, and any derivative mesh output to academic or research purposes only under RAIL Attachment A terms -- not an OSI open-source license, not Apache-2.0 or MIT -- so commercial use isn't permitted despite public GitHub hosting.