Block-Wise Diffusion for Text-to-3D: 5.15x Faster Generation Without Fidelity Loss
Block3D partitions discrete shape-token sequences into contiguous blocks, generating them autoregressively while jointly denoising within each block. On a held-out set from TRELLIS-500K, it reports a 5.15x end-to-end speedup over a fine-tuned autoregressive baseline (25.71s to 4.99s) without sacrificing geometric fidelity. The approach directly targets the computational bottleneck of repeated full-representation processing in global diffusion models.
Block3D introduces a generation paradigm that interleaves autoregressive block generation with joint diffusion denoising within each block. Rather than decoding the entire discrete shape-token sequence in one pass or repeatedly refining a global representation, the method partitions the sequence into contiguous blocks and alternates between predicting a block and denoising its tokens.1
This changes the operative constraint in text-to-3D generation. Autoregressive decoding suffers from sequential error accumulation and cannot revise already emitted tokens; global diffusion and flow models repeatedly process the full representation, making high-quality generation increasingly expensive. Block3D shifts the unit of refinement from the whole sequence to a block, allowing correction of low-confidence tokens within a block before it is finalized.1
The quantitative claim is a mean end-to-end generation time reduction from 25.71 seconds to 4.99 seconds (5.15x) on a held-out set from TRELLIS-500K, relative to a fine-tuned autoregressive baseline, with no reported loss in geometric fidelity.1
The speedup is therefore contingent on the chosen baseline; no comparison to diffusion or flow baselines is reported. We would change our assessment if independent benchmarks show that the block-wise approach falls short of global diffusion models on fidelity, or if the arXiv withdrawal indicates unresolved methodological flaws.1
Context
Text-to-3D methods currently fall into two camps: autoregressive models that decode discrete shape tokens sequentially, and diffusion/flow-matching models that iteratively refine a global 3D representation. The former cannot revise errors, while the latter incurs repeated full-representation processing costs.1
Caveats
The held-out set from TRELLIS-500K is not defined; the hardware and implementation details are not specified; and the method is not compared to existing state-of-the-art diffusion text-to-3D pipelines. These omissions limit the strength of the 5.15x speedup claim. Future code release and evaluation on standard benchmarks (e.g., T3Bench) are needed.1
Computer Science > Computer Vision and Pattern Recognition
Title:Block3D: Efficient Text-to-3D Generation via Block-Wise Diffusion
Abstract:While text-to-3D generation has advanced rapidly, achieving high geometric fidelity at low inference cost remains challenging. Existing text-to-3D methods either decode discrete shape tokens autoregressively or iteratively refine global 3D representations with diffusion or flow models. However, autoregressive decoding is sequential and cannot revise errors, whereas diffusion and flow-matching models repeatedly process the full representation, making high-quality generation increasingly expensive. In this paper, we propose Block3D, a block-wise diffusion framework that partitions the discrete shape-token sequence into contiguous blocks, generates the blocks autoregressively, and jointly denoises all tokens within the current block. To alleviate error accumulation, we introduce confidence-guided intra-block correction, which revises low-confidence tokens before each block is finalized. On a held-out set from TRELLIS-500K, Block3D reduces mean end-to-end generation time from 25.71 seconds to 4.99 seconds, achieving a $5.15\times$ speedup over the fine-tuned autoregressive baseline without sacrificing geometric fidelity.
Submission history
[v1] Thu, 20 Aug 2026 02:12:30 UTC (12,962 KB)
[v2] Fri, 21 Aug 2026 09:59:00 UTC (1 KB) (withdrawn)
[v3] Mon, 24 Aug 2026 01:55:19 UTC (18,547 KB)
[v4] Tue, 25 Aug 2026 05:57:37 UTC (18,546 KB)
Bookmark
Bibliographic and Citation Tools
Code, Data and Media Associated with this Article
Demos
Recommenders and Search Tools
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
The source text above remains the property of its publisher and is reproduced for commentary.
