acceptodds
Under review as a conference paper at ICLR 2027

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 5.15x speedup over the fine-tuned autoregressive baseline without sacrificing geometric fidelity.

Primary area:
generative models
Keywords:
3D Generation, Efficiency, Block Diffusion, text-to-3D

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZeAsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Block3D, a block-wise diffusion framework for efficient text-to-3D generation. Starting from the Cube discrete shape-token representation, Block3D partitions the 1024-token sequence into contiguous blocks, generates blocks autoregressively, and jointly denoises tokens within each active block. It further introduces confidence-guided intra-block correction to correct uncertain codes before a block is committed. On 100 held-out objects from TRELLIS-500K, Block3D improves geometric metrics over Cube while reducing mean generation time from 25.71s to 4.99s.

Critical strengths
  1. The block-wise formulation substantially reduces the sequential decoding horizon, and the reported 5.15× speedup over Cube is impressive.

  2. Block3D and Cube share the same tokenizer, decoder, training subset, and optimization budget, making the comparison relatively well controlled.

  3. The training, decoding, attention-mask, and caching procedures are described in considerable detail.

Critical weaknesses
  1. Limited methodological novelty: The central idea of autoregressive generation across blocks with bidirectional denoising inside each block is highly related to Block Diffusion. Moreover, the M2T/T2T editing mechanism is explicitly adapted from LLaDA2.1. The main novelty therefore appears to be their integration into Cube rather than a fundamentally new 3D generation algorithm. A stronger comparison against a directly adapted Block Diffusion baseline is needed to establish that the proposed 3D-specific components are genuinely necessary.

  2. The main efficiency comparison is somewhat narrow: The headline 5.15× improvement is relative to Cube, whose token-wise autoregressive decoding is inherently sequential. Although other 3D generators are reported, the paper does not provide a controlled block-diffusion baseline using the same Cube representation and backbone. Thus, it is difficult to disentangle the gain from the general block-wise decoding paradigm from the proposed confidence-guided correction and training modifications.

  3. Text-shape alignment does not improve: Block3D improves CD-L1, normal consistency, and F-score, but its CLIPScore is 23.24, lower than Cube's 23.87. Therefore, the claim of improving generation quality should be qualified: the improvement is mainly in geometric reconstruction metrics rather than consistently across all quality dimensions.

  4. Limited evaluation scale and diversity: The main evaluation uses only 100 held-out objects, all sampled from TRELLIS-500K, with one deterministic generation per prompt. This makes the conclusions about general text-to-3D generation and robustness relatively weak. More diverse datasets, multiple random seeds/samples, and statistical significance would make the quality claims substantially stronger.

Questions for authors
  1. Can the authors provide results across multiple random seeds and multiple generated samples per prompt?

  2. Why does Block3D improve geometric metrics while reducing CLIPScore relative to Cube?

Reviewer_VaAbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper improves the generation speed of text-to-3D generation by modifying the generator from dense diffusion model into block-wise diffusion model. The authors model the block-wise relationship as causal relationship so that previous blocks are conditions to the current block's generation. To minimize the accumulation error occured during iterative generation, the author propose confidence-guided intra-block correction to revise low-confidence tokens on the fly. Therefore, the generation speed increase drammatically as well as the quality is improved according to a 100-samples benchmark.

Critical strengths

Technical Approach:

  • Block-wise architecture is a good approach since it can reduce the quadratic complexity of matrix multiplication, which is used widely in the attention layer; therefore decreasing the generation time remarkably.
  • The paper has quite good visualization comparisons.
  • The main paper is well structured, along with detailed experimental settings.
Critical weaknesses

Technical Approach:

  • The block architecture for 3D Diffusion model is not well motivated, there is no theoretical or empirical fundamental evidence that guarantees the performance of the model. This leads to motivation of two following modules are small as well.
  • The shape-code corruption is not defined. It's also not clear how the variable τk\tau_k control the corruption.
  • I am not sure about notations lsl_s, if it's estimated noise at iteration ss then confidence obtained from ls,i,vgl_{s, i, v}^g lacks motivation in Eq.(5). This leads to the confusion when the authors applies the softmax operation on estimated noise later indexed by z^s,i\hat{z}_{s,i}.

Writing:

  • The writing is not good since the paper contains undefined abbreviations and unclear notations. For example, M2T and T2T are not defined in the manuscript.
  • In the concept of classifier-free guidance, given a single conditional encoder, we can obtain the positive and negative noise conditioned on positive and negative prompts. However, in this paper, the author denotes two encoder C+C^+ and C−C^-, which I think there is a mistake in expression here.
  • τk\tau_k is not clearly defined the controlability on the shape corruption.

Experiments:

  • The three metrics CD-L1, NC, and F@1% might be used wrongly since given a text prompt, there can be a large number of variations that are related to that prompt, which gives no guarantee that the model can reconstruct correctly the expected mesh.
  • 100 samples for evaluation is quite small, the method should be evaluated with some external benchmark [1][2]

[1] T3Bench: Benchmarking Current Progress in Text-to-3D Generation. [2] GPT-4V(ision) is a Human-Aligned Evaluator for Text-to-3D Generation.

Questions for authors

Please address my weaknesses as well. I hope to receive explaination from authors.

Q1: Can the authors provide empirical obversations, evidence that motivate the use of block-based diffusion model? Q2: Can the author provide empirical observations when using block diffusion model that leads to the need for confidence-guide decoding? Q3: Can the author explain the mathematical motivation for applying argmax and softmax operations on the estimated nosie ls,i,vgl_{s,i,v}^g? Q4: The authors did not mention the definition of τk\tau_k, how it can control the corruption. Q5: What is the shape corruption? Q6: Which orders are the blocks sorted for causal generation?

Reviewer_DSP1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Block3D is a block-wise diffusion framework for text-to-3D generation that addresses the long-standing trade-off between geometric fidelity and inference cost. Block3D shifts the causal dependency from individual shape tokens to contiguous latent blocks: blocks are generated left-to-right, but all tokens within the currently active block are jointly denoised in parallel under bidirectional attention. Each committed block is then frozen, cached, and used as a prefix for the next block, substantially shortening the sequential decoding horizon without changing the underlying tokenizer or mesh decoder. On a held-out set of 100 TRELLIS-500K objects, Block3D achieves the best geometry metrics among evaluated methods while it is the fastest system tested, with mean end-to-end generation time of 4.99 seconds versus 25.71 seconds for the fine-tuned Cube baseline.

Critical strengths
  • Efficiency gain: Block3D cuts mean end-to-end generation time from 25.71s (fine-tuned Cube baseline) to 4.99s

  • No quality sacrifice, and actually better geometry. It achieves the best geometric fidelity among compared methods

Critical weaknesses
  • Errors can never be fixed once a block is committed. All revision is intra-block; completed prefix blocks are frozen, and the paper explicitly states it does not remove cross-block exposure bias.

  • Strong sensitivity to hyperparameters with little headroom. Larger blocks trade quality away quickly (F@1% drops from 0.309 at B=64 to 0.103 at B=256), and the reported speedup relies on a specific tuned configuration (B=64, T=4, thresholds 0.95/0.9, guidance 3.0).

Questions for authors

Hyperparameter sensitivity: almost all hyperparameters (e.g., linear guidance) are all fixed with no reported ablation. How were they chosen, and how sensitive are the results to each?

Rather than comparing against independently configured baselines, it would strengthen the paper to compare against Cube and TRELLIS at matched latency (and matched quality) — e.g., a Pareto plot of F@1% vs. seconds across all methods and your own block-size/steps ablations.

Diversity and failure modes are unreported. You generate exactly one shape per prompt and evaluate against a single ground-truth mesh. Have you measured sample diversity across multiple draws from the same prompt (necessary given deterministic argmax decoding), and what is the rate of invalid/degenerate outputs? Appendix C.5 mentions invalid outputs, but nothing appears in the main paper.

1 comment

Thank you for the response. I'll keep the score unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.