Block3D proves 5x faster 3D mesh generation doesn't have to sacrifice fidelity

The core problem Block3D addresses is that generating high-fidelity 3D meshes is slow whether you use autoregressive token-by-token decoding or full-sequence diffusion. Token-wise autoregression is sequential by design; full-sequence denoising revisits every position on every pass. Block3D breaks this tradeoff by shifting the causal unit from individual shape codes to contiguous blocks. Blocks are generated left to right, but all positions within the active block are denoised jointly under bidirectional attention. Once a block is committed, it is frozen and becomes part of the prefix for subsequent blocks.

The method builds directly on Cube's frozen vector-quantized shape representation, which encodes each mesh as a fixed-length sequence of 1,024 discrete codes. Block3D replaces only Cube's token-wise prior with a block-causal generator, leaving the tokenizer, codebook, CLIP text encoder, and mesh decoder untouched. This makes the controlled comparison clean: both models share the same initialization, geometry components, training data, and optimization budget, differing only in how the prior generates the code sequence.

The confidence-guided correction mechanism is where the design gets interesting. Within each active block, two complementary update operations run before the block is finalized: mask-to-token filling for still-masked positions, and token-to-token replacement for already-filled positions whose new candidate exceeds a confidence threshold. A deterministic reveal quota guarantees every block completes within the fixed denoising horizon regardless of confidence scores. The ablation is direct: adding token-to-token editing over mask-only recovery reduces Chamfer distance by about 4.7% and raises F-score at the 1% threshold from 0.287 to 0.309.

The speed result is the headline number. On 100 held-out objects from TRELLIS-500K, Block3D's mean end-to-end generation time is 4.99 seconds versus 25.71 seconds for the Cube baseline, measured on a single A100. The paper is careful about what that count includes: condition encoding, shape-code generation, and mesh decoding are all timed; model loading and disk I/O are excluded. Median and 90th-percentile latency improve consistently, so the gain is not driven by a handful of easy prompts.

The geometry results are equally notable. Block3D achieves the best Chamfer distance, normal consistency, and F-score among the evaluated methods, which include TRELLIS-text, AR3D-R1, and ShapeLLM-Omni. CLIPScore stays competitive. The block-size ablation shows the quality-speed tradeoff clearly: a block size of 96 cuts mean latency to 3.62 seconds but drops F-score to 0.279, while a block size of 64 holds F-score at 0.309 at 4.99 seconds.

The paper is honest about what the method does not fix. Completed blocks are immutable, so cross-block exposure bias remains. The confidence-gated correction only operates within the active block before commitment; errors that propagate across block boundaries cannot be revisited. The authors flag cross-block refinement as future work. The evaluation is also bounded: 100 paired prompts from one dataset, one generation per prompt, one GPU. That scope is enough to establish the speed-quality tradeoff but leaves generalization to broader shape distributions open.