acceptodds
Under review as a conference paper at ICLR 2027

Blitz: Accelerating the Diffusion Module of AlphaFold 3 and Boltz-2

Abstract

Diffusion-based protein structure prediction models produce all-atom structures in about 200 denoising steps, and that cost has driven a number of few-step samplers. We present Blitz, a few-step distillation framework that uses a geometric noise schedule to distribute steps more evenly across noise scales, then learns those updates through multistep consistency training. Physical steering and target-free candidate selection help reduce the remaining geometric defects at inference. We apply Blitz to AlphaFold 3 and Boltz-2, reducing denoising from 200 steps to eight. Because published few-step results use different benchmarks and scoring procedures, we compare both distilled models with their teachers and other few-step methods under a common evaluation protocol covering structural accuracy, interface quality, ligand placement, and stereochemical validity. Blitz retains teacher-level structural accuracy, with AF3-Blitz outperforming all evaluated few-step baselines and exceeding its teacher on several accuracy metrics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein structure prediction, diffusion models, distillation, consistency models, structural biology

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NbUROpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Blitz accelerates AlphaFold 3 and Boltz-2 by reducing diffusion sampling from 200 steps to just 8, while largely preserving structural accuracy and molecular geometry.The key observation is that simply reducing diffusion steps creates two problems 1)structural accuracy deteriorates and fewer denoising steps can produce less accurate structures.2)Molecular geometry deterioration.Even when predicted protein folds look correct, structures can contain severe atomic clashes, incorrect bond lengths, or distorted bond angles.The authors argue that previous accelerated folding methods have underestimated the second problem because conventional evaluation metrics do not always penalize these geometric errors.

Critical strengths
  1. The authors show that AF3's original diffusion schedule is poorly suited for few-step sampling. Simply changing the schedule improves structural accuracy and molecular geometry without retraining.This is arguably their strongest technical contribution (Figures 2–3).
  2. They demonstrate that high lDDT can conceal severe geometric defects.This exposes a meaningful weakness in conventional benchmarking.
  3. Strong AF3 distillation performance is shown.AF3-Blitz reduces diffusion from 200 to 8 steps while maintaining approximately the same structural accuracy.The result is technically impressive, even if the methods are relatively established.
Critical weaknesses
  1. Limited methodological novelty.The method combines existing techniques:geometric noise scheduling from EDM,multistep consistency distillation,physical steering and Geometry-based candidate selection. There is no fundamentally new distillation framework or architecture. Can the authors clearly reveal what is novel for clarity?
  2. Missing comparison with DCFold.Despite discussing DCFold extensively, the authors do not compare against the trained model. This weakens the claim of superiority over previous few-step approaches. DCFold seems to have not a repo. can you compare in-paper results?
  3. No end-to-end computational efficiency analysis.The headline 25× reduction concerns diffusion evaluations, not overall inference time. Without wall-clock measurements, the practical benefit is unclear. Can you include end-to-end budget analysis?
  4. Incomplete coverage of newer models.Protenix-v2 and OpenDDE is excluded, and the framework is demonstrated primarily on AF3 and Boltz-2.
  5. Quality preservation is not universal...For Boltz-2, eight-step Blitz reduces ligand success from 36.5% to 30.6% and PB-validity from 60.9% to 52.0%.
Questions for authors

1.The proposed method combines geometric noise scheduling, multistep consistency distillation, and physical steering, all of which have been previously introduced. What is the fundamental methodological contribution beyond adapting and combining these existing techniques for AF3 and Boltz-2? In particular, which component is essential to the reported improvement? Provide detailed ablations

2.Missing DCFold comparison DCFold is one of the closest related approaches, and the manuscript discusses its limitations extensively. However, the experiments include only a related untrained ODE baseline rather than the trained DCFold model. Why is DCFold not included in the experimental comparisons? If code is not available, please include paper-reported values and compare. Can the authors compare Blitz and DCFold under matched teacher models, diffusion budgets, and evaluation protocols?

3.Does reported 25× acceleration applies only to diffusion evaluations?.Pairformer/trunk computation can represent a substantial fraction of total inference time.What is the actual end-to-end inference speedup achieved by Blitz? Could the authors provide wall-clock runtime and GPU-cost comparisons across different protein sizes, including trunk computation, multiple candidates, and candidate selection? This is particularly important for determining whether Blitz provides a meaningful advantage in practical structure prediction pipelines.

4.Generalization to Protenix-v2 and OpenDDE .Does the geometric noise schedule and consistency-distillation approach generalize to Protenix-v2?Given that the proposed improvements depend on the characteristics of the pretrained diffusion model, it would be useful to establish whether the conclusions hold for newer architectures.

5.Contribution of distillation versus candidate selection The paper reports improvements in stereochemical validity, but Appendix G.3 shows that geometry-aware candidate selection substantially reduces violation rates.For example, AF3-Blitz (16) improves from approximately 13.9% to 6.9% of targets containing violations after reranking.How much of the reported improvement comes from the distilled diffusion model itself versus the geometry-based reranking procedure?Does Blitz still outperform competing methods when all models use identical candidate selection procedures?

  1. Preserving biologically meaningful accuracy Although Blitz largely preserves global structural accuracy, some important metrics deteriorate. For example, Boltz-2-Blitz (8) reduces ligand success from 36.5% to 30.6%.Does aggressive diffusion distillation disproportionately affect difficult molecular interactions, such as antibody–antigen interfaces or protein–ligand binding poses?Could the authors provide stratified evaluations of interface accuracy, particularly for challenging complexes where the teacher is already uncertain?

7.Could the authors provide code? Reviewer would like to double-check

2 comments
AuthorsOpenReview ↗

W1 & W2: On Novelty & DCFold

On Novelty: We thank the reviewer for their time and for highlighting the strengths and weaknesses. We believe the questions raised have helped make this manuscript better. We have added an additional section (Section 5.2) in the paper that helps separate the advantages of the schedule, sampling, reranking, and training. However, on the question of limited novelty, we respectfully disagree. First, we challenge the prevailing use of the polynomial schedule introduced by Karras et al. (2022), which has become the de facto schedule in almost all structure prediction implementations because of its favorable properties. Our finding that geometric spacing benefits the few-step prediction regime of co-folding models is new. In image generation, for example, we observe the opposite: FID, the metric used to assess generation quality, worsens with this schedule. Additionally, the schedule and sampling choices interact in ways that our experiments help tease apart and is a new methodological insight.

8-step scheduleMean FID
Polynomial (ρ = 7)15.666
Geometric17.993

ImageNet-64, using 50,000 generated images (lower is better)

Existing work on accelerating diffusion modules either applies only to specific settings, such as ligand-pose prediction (DeCAF), or is not reproducible (see commentary on DCFold below). Techniques such as consistency distillation alone or flow maps that have been used prior to this work have not matched teacher level fidelity and multistep consistency distillation (MSCD) resolves this gap with our schedule enhancements. Our initial experiments with alternative mechanisms like progressive distillation and multistep moment matching did not transfer to the co-folding regime. Combining our choices of schedule, sampling, multi-step consistency training, and re-ranking to accelerate prediction without sacrificing accuracy & geometry is a nontrivial contribution to open science. We believe all of this aligns with ICLR’s scope of contributing new knowledge and value to the community.

On DCFold: We corresponded extensively with the authors to reproduce their reported results but were unable to do so. They provided a code snippet from their training algorithm, but our model consistently collapsed after approximately 1000 steps using their proposed TGM sampling method. The authors acknowledged that this was something they observed as well and suspected training to be sensitive. They expected that not every run would reliably produce a well performing model. We also requested outputs from their trained model on the PDB 384 benchmark reported in their paper. However, the DCFold authors declined to share them, citing licensing constraints. Our main concern with the paper is illustrated in Table 1. Here, the untrained one-step Protenix ODE baseline used in their experiments has similar unchecked lDDT to its teacher on a sufficiently large amount of targets. Therefore, we believe a blanket comparison with the reported result would paint the wrong picture.

Reference

Karras, T., Aittala, M., Aila, T., and Laine, S. (2022). Elucidating the Design Space of Diffusion-Based Generative Models. Advances in Neural Information Processing Systems, 35.

AuthorsOpenReview ↗

W3, W4 & W5: Speedups, Newer Models and Quality

On Timing: We agree with the comment on timing and have added a table to the appendix detailing inference speedups for AlphaFold3 separating the trunk, diffusion and confidence pipeline (Appendix A.3, Table 7). Diffusion dominates inference time for proteins with up to 512 amino acids, and our sampler reduces diffusion time to a few hundred milliseconds. Complementary methods optimize the trunk through faster attention mechanisms or kernel optimization (See PairMixer and TurboPairFormer). These could be incorporated into our pipeline, but we deliberately leave trunk optimization outside the scope of this work. Candidate selection is a CPU only task and is negligible.

On Newer Models We have also added the controls for Proteinx v2 and OpenDDE with just our schedule, sampling and ranking changes showing significant improvement (Appendix A.4, Table 8). However, retraining them with MSCD requires significant compute and is out of scope for this paper.

Selected metrics from Appendix A.4, Table 8. K8 uses our geometric schedule and exponential sampler; all configurations use the original teacher weights.

MetricProtenix-v2: stock (200 steps)Protenix-v2: K8OpenDDE: stock (200 steps)OpenDDE: K8
Complex lDDT ↑0.85890.85980.85560.8514
PB-valid (%) ↑45.742.739.831.5
lDDT-PLI ↑0.63770.63540.62530.6171
Bonds / target ↓0.211.740.221.50
Clashes / target ↓5.1216.016.4118.51
Any violation (%) ↓45.353.718.554.6

Candidate Selection: The additional section on 5.2 helps clarify and compare the differences of candidate selection. Blitz outperforms competing methods is true regardless of the re-ranking procedure. However, the goal is teacher-level parity and for that, in extensive experiments and across benchmarks, we notice that the geometric reordering is an important mechanism. We report geometry-ranked competing methods in Appendix A.5 (Table 9) and native-confidence-ranked Blitz configurations in Appendix H.3 (Table 15).

On Quality: We agree that global structural accuracy alone does not establish preservation of biologically meaningful interactions. We have added a confidence-stratified protein–ligand evaluation in Appendix A.6, Table 10. Both teachers and students perform worse on low-confidence complexes. AF3-Blitz largely maintains or improves ligand-interface accuracy across the difference confidence levels. For Boltz-2, the additional loss could be a contribution of teacher limitations and architectural choices which are difficult to separate. However, the 16-step variant is within ~2.6 pp of that level and can give users a practical choice between the compute they want to allocate for this particular task. This analysis addresses protein–ligand interactions but not antibody–antigen interfaces.

AF3 ligand performance stratified by teacher confidence (Appendix A.6, Table 10).

MetricModelLowMediumHigh
lDDT-PLI ↑Teacher0.48410.61810.8222
Blitz K80.48290.61420.8234
Blitz K160.48170.61640.8260
Primary < 2 Å (%) ↑Teacher48.262.088.4
Blitz K846.163.788.4
Blitz K1647.563.488.5
PB-valid (%) ↑Teacher28.035.866.1
Blitz K823.335.464.7
Blitz K1625.939.066.7
Success (%) ↑Teacher14.421.462.0
Blitz K812.726.561.7
Blitz K1614.525.761.3

Q7 We have provided the training and inference code as supplementary material. The model weights of Boltz-2 Blitz will be made public upon acceptance. Due to licensing restrictions by DeepMind, AlphaFold3 Blitz will be excluded from this release.

References

Ouyang-Zhang, J., et al. (2026). Triangle Multiplication Is All You Need For Biomolecular Structure Representations. International Conference on Learning Representations (ICLR).

Ran, Y., et al. (2026). TurboPairFormer: Fast and Stable Protein Folding Model Training with an Optimized Triangle Attention Kernel. arXiv:2610.05854.

Reviewer_Dpu6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Geometric scheduling and the corresponding distillation scheme, as well as geometric reordering. Finally, it implements AF3 with only a few steps.

Critical strengths
  1. This article examines the issue of geometric validity, a crucial point in the field of protein synthesis.
  2. This work involved many experiments, including sensitivity analysis, etc.
Critical weaknesses
  1. This work lacks a comparison between the original weights and the distillation weights under the same geometric scheduling and full inference configuration. This experiment is crucial for the additional advantage of separating distillation training. Without this experiment, I would have no way of determining whether distillation or geometric scheduling is responsible for the effect.

  2. When explaining baseline comparisons, it's necessary to determine whether the baselines used geometric reordering or were averaged.

  3. Missing important baseline Original AF3/boltz-2 weights + geometric 8 steps + same update formula, step size coefficient and geometric reordering

  4. For protenix-mini, you need to use either the 2-step method described in the paper or the 5-step method from the open-source software, instead of the 1-step method. Please compare again.

Questions for authors
  1. The current experiments cannot attribute the results of the paper to geometric scheduling, distillation, or geometric reordering, especially since the effectiveness of distillation has not yet been proven.

  2. If all four experiments mentioned above are completed, I can consider raising the score. Even if distillation is not very useful, I still think that geometric scheduling itself is a good contribution.

1 comment
AuthorsOpenReview ↗

Separating Numerical Design Choices

We thank the reviewer for the positive assessment, and we strongly agree that the suggested controls are very important. To that end, we have now added a new section (Sec. 5.2) after the main results to demonstrate how much each individual component contributes, including the noise schedule, sampling, reranking, and training. The change in noise schedule alone of the teacher weights is also added to Table 2. Apart from accuracy, we would especially like to draw attention to the last row, which shows that the geometric schedule with our exponential sampler has the fewest structures with violations (without any training!), at 37.1% of the approximately 2,300 structures. MSCD further helps reduce that down and improves accuracy to reach teacher level fidelity (also much lower violations in the 16-step model). We have also made the reranking configuration explicit either under the footer or in each column of all tables. Table 3 reports the competing methods using their native ranking. Table 9 applies geometric ranking to the competing methods discussed. Table 15 reports Blitz with native-confidence ranking.

We apologize for the error in the reported sampling budget. The original Proteinx Mini experiment used 5 NFEs, and the reported value was a transcription error. To confirm, we have rerun the experiments with both NFE = 2 and NFE = 5 configurations of the Proteinx Mini model and added the results to Table 3.

AF3 at 8 steps and 3 recycles — controls from Table 4 and trained students from Table 2

MetricPolynomialPolynomialPolynomialGeometricGeometricGeometricPolynomial studentBlitz K8
SamplerEuler, η=1\eta=1Exponential, Eq. (7)Exponential, Eq. (7)Euler, η=1\eta=1Exponential, Eq. (7)Exponential, Eq. (7)Exponential, Eq. (7)Exponential, Eq. (7)
WeightsTeacherTeacherTeacherTeacherTeacherTeacherTrainedTrained
SelectionGeometricConfidenceGeometricGeometricConfidenceGeometricGeometricGeometric
Complex lDDT ↑0.77710.82030.82780.82550.85210.85640.86030.8635
lDDT-PLI ↑0.61160.63840.63560.60960.64020.63660.63940.6425
PB-valid (%) ↑25.428.425.535.439.437.035.443.1
RF-valid (%) ↑34.437.235.935.439.438.736.440.5
Bonds / target ↓117.7068.4651.2649.0435.8925.122.830.79
Angles / target ↓54.2233.1522.9720.7113.798.711.010.23
Clashes / target ↓67.82149.1789.1738.42161.2587.494.013.80
Any violation (%) ↓99.099.097.583.164.237.150.511.3
Reviewer_VrJaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work identifies that prior AF3-like distillation models have largely neglected physical-plausibility metrics when evaluating generated molecular structures. To address this gap, the authors propose a method that combines multi-step consistency distillation, a geometric noise schedule, and physical steering.

Critical strengths

The proposed method improves the physical plausibility of generated structures relative to previous AF3-like distillation work, (but only at N_cycle = 3).

Critical weaknesses

The evaluation is concentrated on N_cycle = 3, a regime that is not representative of practical deployment settings. Without results at more standard configurations such as N_cycle = 10, it is unclear whether the reported gains persist under realistic inference budgets, which weakens the practical impact of the method.

Questions for authors
  1. Could you add results at N_cycle = 10 to show that the improvements in physical plausibility hold beyond the N_cycle = 3 regime?
    2. Suggestion: A short background section on multi-step consistency / consistency distillation would make the technical motivation more accessible to readers less familiar with that literature.
1 comment
AuthorsOpenReview ↗

Ablations on Recycling

We thank the reviewer for their positive assessment. We agree that the recycle control is an important variable and we have now added additional ablations to control for it. We find that all of our previous results still hold. We’ve added this to a new section on numerical design (see Sec. 5.2 Table 4), where the last three columns compare the teacher and both students with 10 recycles. The extended table is provided in Appendix A.2. We’ve also added text in the background section to provide an intuitive comparison between consistency models and their multistep variants. Supporting this, we’ve also added a new figure (Figure 4) beside our algorithm to make it clearer to readers what the model is learning.

AF3 at ten recycles (from Table 4)

MetricTeacherBlitz K8Blitz K16
SelectionNative confidenceGeometricGeometric
NFE200816
Complex lDDT ↑0.86240.86360.8639
lDDT-PLI ↑0.64200.64750.6487
PB-valid (%) ↑44.644.946.6
RF-valid (%) ↑42.540.642.7
Bonds / target ↓0.060.820.13
Angles / target ↓0.090.260.09
Clashes / target ↓0.663.711.79
Any violation (%) ↓9.911.36.8

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.