acceptodds
Under review as a conference paper at ICLR 2027

PatchAFR: Noise-Corrected Patch-Adaptive Frequency Routing for Pixel-Space Diffusion Models

Abstract

Pixel space diffusion models generate images directly from raw pixels without relying on pretrained tokenizers, but their high dimensional input space makes training from scratch expensive. Existing frequency based approaches mainly characterize frequency recoverability at the global image level. We observe that this global treatment overlooks an important property of images: the same frequency band can have different local signal strength relative to the diffusion noise across spatial regions, resulting in spatially varying recoverability. Based on this observation, we introduce PatchAFR, a lightweight **Patch**-**A**daptive **F**requency **R**outing method that can be integrated into pixel space diffusion models. We formulate frequency recoverability through local spectral evidence and derive a timestep aware correction to estimate this evidence from noisy inputs. PatchAFR then jointly uses the corrected local evidence and diffusion timestep to adaptively route frequency bands for each spatial patch. Experiments on the pixel space diffusion model JiT demonstrate improvements throughout training. On JiT B/32, PatchAFR reduces FID by 8.4%, 7.7%, and 9.1% at epochs 60, 120, and 200, respectively, and requires 29% fewer training epochs to reach the same FID over 600 epochs.

Primary area:
generative models
Keywords:
pixel-space diffusion model, frequency-aware learning, patch-adaptive routing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_5DF9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends spectral forcing by changing the high-frequency filtering to depend on the patch, rather than having one global cut off. For this, PatchAFR splits the image into 4 DCT freq bands, and for evey patch a tiny MLP decides how much of each band should pass through. To showcase the approach, experiments are conducted on ImageNet, using primarily a JiT-B/32 model.

Critical strengths
  • The method is simple and parameters efficient, acting a plug-in, making it potentially easy to adopt.
  • Good improvements for the B/32 model.
  • Reasonable number of ablations where conducted for some of the main components.
Critical weaknesses
  • Novelty is limited and the related work is incomplete. The method appear to be essentially SF+a learned per patch soft gate over the DCT bands, with the gate budget taken direct from the SF's mask. Most of this components were used in a similar manner before and the paper doesn't currently explain how they differ. E.g: [1] Frequency-Aware Flow Matching for High-Quality Image Generation, CVPR 2026 (they use a similar timestep-cindioned gating of freq bands), [2] Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation, CVPR 2026 and [3] Diffusion Models With Learned Adaptive Noise, NeurIPS 2025 (both establishing the same premise that denoising difficulty varies spatially).

  • The theory its rather standard and is classical spectral substraction. The unbiasedness and variance results from appendix R are also textbook moments of a Gaussian quadtratic form (see Mathay&Provost, "Quadratic forms in Random Variables: Theory and Application"). The derivations look correct anyway, but the paper should present this parts (including in the main) as background and as a theoretical grounding/contribution.

  • The claims of general improvements are overstated, especially that the scaling appears to be poor beyound B/32 models. In the standard JiT setting, PatchAFR is worse than the JiT baseline at 120 and 200 epochs. So it would appear that both under a longer scheduler and using larger models the approach doesn't offer any notable gains. This is particularly important because the B/32 setting at 256px is rather constricted configuration.

  • Some baselines and evaluations protocols make the results harder to interpret and may be unfair/overclaim: For example, herein all methods use a cfg of 4.0 while JiT paper tunes the CFG per model. As a results the JiT paper reports 4.37 for the B/16 model at 200 epochs vs only 4.77 in this work.

  • Continuing on the same line as above, other methods that speed up JiT training are not compared at matched epochs. WiT: Waypoint Diffusion Transformers via Trajectory Conflict Navigation is cited but not compared, and appears to be better at 200 epochs.

  • PixelREPA is not reported standalone. The work reports 4.14 for B/16 + PixelREPA which is bellow PixelREPA standalone number of 4.0 in the asme setting.

  • All experiments are done in small scale and constrained setting, unclear if they can scale and current evidence points in the opposite direction.

Questions for authors
  • Because N_pb is a global band projector averaged over a patch, it is nearly constant across patches, minus the boundary effects. Is in this case the "noise correction" close to a deterministic function which the MLP could in principle learn on its own since it already receives t?

  • Can the paper explain how the out of domain baselines are adapted?

  • What would be the accuracy under random, shuffled or fixed-per-band gates under the same budget? Without these, the gains could come from a different global attenuation shape (i.e. soft per-band gate vs SF's sigmoid cutoff).

2 comments
AuthorsOpenReview ↗

Part I: W1-W3

Dear Reviewer 5DF9,

We sincerely thank you for your positive assessment of our work and for your questions.

W1: Novelty and the related work.

  • Novelty. We would like to respectfully offer a slightly different perspective on the novelty of our work and clarify our contributions. Our work introduces soft gates for frequency-domain modulation, but it is not simply a combination of soft gating and Spectral Forcing (SF). Frequency routers have been used in image processing and generation for many years. We clarify the novelty of PatchAFR and its differences from SF in two aspects:
  1. PatchAFR learns different routing weights for frequency bands in each patch. SF uses a timestep-dependent DCT low-pass mask for frequency filtering. PatchAFR uses only the mean value of SF mask as a scalar budget prior, while the learned router determines the frequency allocation for each patch.

  2. Raw frequency energy measured from noisy pixel inputs contains diffusion noise. PatchAFR uses the known diffusion noise contribution to construct routing features. Most previous work does not model the noisy inputs of diffusion models.

  • Related work. We thank the reviewer for pointing out the missing related work. We will add these discussions to the Related Work section in the revised manuscript.

FreqFlow[1] and our method both introduce frequency-aware modules, but their architectures and roles are different. [1] employs a frequency branch to fuse features with spatial branch. In contrast, PatchAFR is a lightweight input routing plug-in. Its router predicts per-patch gates for four DCT frequency bands of the noisy input, directly adjusts the contribution of each band, and passes the recombined input to the original backbone. Therefore, the frequency weights in [1] are used for feature fusion, while PatchAFR's gates are used for input frequency routing without introducing an additional frequency modeling branch.

PatchForcing[2] also considers differences in denoising difficulty across patches, but mainly controls the denoising process by adjusting patch-wise denoising timesteps. PatchAFR instead focuses on how information is allocated across different frequency bands for each patch.

MuLAN[3] learns a data-dependent noise process to adjust how noise is injected, while PatchAFR learns how to allocate and use information from different frequency bands given a noisy state.

W2: The theory is standard

General related statistical theory about noise and energy has been well explored in prior work. Indeed, the purpose of the theory in our current manuscript is to motivate the noise-corrected routing features and as a theoretical background. We will polish wording for proper statement.

W3 & W7: scaling about training, model size and patch size.

We appreciate your insightful observation, which closely aligns with our experimental settings and findings presented in Table 2 and Figure 4 of our manuscript.

  • B/16. We provide a detailed analysis of this behavior in Budget Sensitivity on B/16 in the appendix. Following your questions about eval protocols and PixelREPA, the latest experiments show our approach can further improve PixelREPA and achieve state-of-the-art results on epoch100 of B/16.
EpochPixelREPA (Paper)PixelREPA (Reproduced)PatchAFR + PixelREPA
1005.95.925.09
2004.0N/A4.00
  • Model scaling. We observe consistent performance improvements not only on JiT-B/32 but also on the larger JiT-L/32. As shown in Table 2, PatchAFR outperforms the corresponding JiT baselines on both B/32 and L/32 at all reported training stages (60, 120, and 200 epochs).

  • Training scaling. Figure 3 presents the training results on B/32 up to 600 epochs. PatchAFR maintains its performance advantage throughout longer training and requires 29% fewer training epochs to reach the same FID.

  • Patch 32 settings are important. Large-patch modeling is a widely studied direction for efficient pixel-space generation. Recent works such as MiniT2I, PixelU, etc. use patch size 32 in experiments. Notably, PixelDiT also experiments with B/32 at 256×256 resolution. Larger patch sizes reduce the number of tokens and the computational cost of Transformers. Therefore, improving generation quality and training convergence for large-patch pixel diffusion has practical value.

AuthorsOpenReview ↗

Part II: W4-Q3

W4: evaluation protocols.

Thank you for pointing it out. We use CFG=4.0 consistently in the manuscript to directly compare different methods under the same sampling conditions and avoid the influence of different CFG settings on the comparison. Evaluation results across different CFG scales and random seeds are also provided in the appendix.

Following your suggestion and the CFG tuning protocol in the original JiT paper, we further searched for the best CFG separately for each model. For B/32 and L/32, the tuned results are consistent with the performance reported in the manuscript. For B/16, the best CFG and corresponding FID-50K are shown below:

ModelBest CFGFID ↓
JiT-B/16 (reproduced)3.54.49
PatchAFR-B/163.24.72
PatchAFR-B/16 + PixelREPA3.54.00

Compared with the results reported in the manuscript using a fixed CFG=4.0, CFG tuning improves the FID of JiT-B/16, PatchAFR-B/16, and PatchAFR-B/16 + PixelREPA from 4.77, 5.12, and 4.14 to 4.49, 4.72, and 4.00, respectively. We will include these evaluation results in the revised manuscript and clearly distinguish between fixed-CFG comparisons and best-CFG evaluations.

W5: Other methods compared at matched epochs.

WiT reports its results on JiT-B/32 using ImageNet 512×512, while our experiments use ImageNet 256×256.Therefore, FID results at the same epoch are not directly comparable due to the different experimental settings.

W6: PixelREPA standalone.

We thank the reviewer for highlighting this important baseline. We have added experiments with JiT-B/16 + PixelREPA up to epoch 100 due to time and resource limitations.

On B/16, our method combined with PixelREPA achieves a large improvement over PixelREPA at epoch 100, meeting our goal of accelerating training.

The early-stage convergence improvement observed on vanilla JiT (Table 2) also extends to the combination with PixelREPA. At epoch 100, PatchAFR + PixelREPA reduces FID from 5.92 to 5.09, while matching PixelREPA's reported FID of 4.00 at epoch 200. These results suggest that PatchAFR complements PixelREPA by further improving training efficiency without sacrificing its later-stage performance.

EpochJiT + PixelREPA (paper)JiT + PixelREPA (our reproduction)Ours + PixelREPA
1005.95.915.09
2004.0N/A4.00

Q1: effectiveness of N_pb

That’s an insightful question! Mathematically, (Np,bN_{p,b}) is indeed deterministic. Ideally, the router may be able to learn this transformation. In practice, there is a trade-off between model capacity and overhead. A lightweight MLP has limited learning capacity. Noise correction provides an inductive bias rather than additional information. By explicitly subtracting the expected noise floor, we use the known relationship from the forward diffusion process. This means the MLP does not need to learn this relationship implicitly through optimization alone, helping it improve its capability more quickly.

As shown in Table 5, the MLP using raw energy does not reach the performance of explicit noise correction in the middle and later training stages. This shows that explicit guidance is still beneficial.

Q2: details about the out of domain baselines

Thank you for your question. We have updated the implementation details of out-of-domain in the Appendix C “Adaptation of Frequency-Based Baselines.”

Q3: accuracy under random, shuffled or fixed-per-band gates under the same budget?

Relation and differences from SF. Spectral Forcing (SF) applies a timestep-dependent DCT low-pass mask, while PatchAFR uses its mean value only as a budget prior. Our learned router allocates this budget across patches and frequency bands.

Following your suggestion, we conducted budget-matched ablations with random, shuffled, and fixed-per-band gates (Appendix A, The Router Effectively Learns Frequency Allocation). Their FID-50K scores are 229, 249, and 243, respectively (CFG=4.0), with no high-quality samples observed. These results show that the SF-derived budget alone is insufficient, highlighting the importance of learned frequency allocation.

We sincerely appreciate your insightful and constructive review! Following your suggested experiments and evaluation protocols, we have significantly improved our work, obtained more comprehensive results, and built stronger connections with related studies. More excitingly, we have also uncovered previously underexplored benefits of our method, showing the great potential of B/16 when combined with PixelREPA and achieving even better results. We are truly grateful for your thoughtful feedback and suggestions, which have helped us make meaningful improvements to our work.

Reviewer_WQB8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PatchAFR, a lightweight, patch-adaptive frequency routing module for pixel-space diffusion models (PDMs) that operates directly on the noisy input. The key idea is to keep a global DCT-based frequency decomposition but make routing decisions locally per patch by estimating noise-corrected local spectral evidence and learning a small MLP router that gates band contributions under a timestep-aware frequency budget. Integrated with JiT on ImageNet 256×256, PatchAFR yields consistent FID improvements across checkpoints (e.g., 9.1% at epoch 200 on JiT-B/32) and achieves the same FID with 29% fewer training epochs, with minimal parameter and compute overhead.

Critical strengths
  1. The paper clearly motivates the problem: even when frequency bands are defined globally, the recoverability of their contents varies across spatial locations. It connects this motivation to a concise formal formulation.

  2. Under the large-patch configuration, the method delivers consistent performance improvements with very few additional parameters. On JiT-B/32, PatchAFR outperforms both JiT and Spectral Forcing at all three reported training checkpoints.

Critical weaknesses
  1. The experiments primarily focus on large-patch configurations, with limited investigation of the more commonly used 16×16 patch size. Moreover, PatchAFR does not demonstrate consistent improvements with 16×16 patches: it underperforms the JiT baseline after 60 epochs, and PatchAFR + PixelREPA also underperforms JiT-B/16 + PixelREPA at epoch 200.

  2. Evaluation across different backbones is also limited. The method is not tested at 512×512 resolution or on other Pixel Diffusion backbones, such as Deco and PixelDiT. These architectures lack a CNN bottleneck and may therefore be more sensitive to frequency modulation, potentially making the benefits of the proposed method more apparent.

Questions for authors
  1. Is it necessary to compute the high-frequency components in Band 3? How would removing this band or reducing its weight affect visual quality and quantitative results? Given the high dimensionality of images, the highest-frequency components may have a relatively limited impact on visual perception.

  2. Training was stopped relatively early at 200 epochs, making it difficult to assess the results after full convergence. Moreover, the experiments mainly focused on the B-sized model. Would using more epochs, such as 600 or 800 epochs, or testing other model sizes, still be effective?

2 comments

Dear Reviewer WQB8,

We sincerely appreciate your positive assessment of our work and your thoughtful questions.

W1. Performance on 16×16 patches.

Thank you for raising this important point! We have partially discussed this issue in the original manuscript about budget prior and have conducted additional experiments about JiT-B/16 + PixelREPA to further investigate it.

  • We further conducted experiments on JiT-B/16 + PixelREPA. Due to time and computational resource constraints, we reproduced the PixelREPA baseline up to epoch 100. Following Reviewer 5DF9’s suggestions and the CFG search procedure used for FID reporting in JiT, we report the results below rather than evaluating at a fixed CFG scale of 4.0. The early-stage convergence improvement observed on vanilla JiT (Table 2) also extends to the combination with PixelREPA. At epoch 100, PatchAFR + PixelREPA reduces FID from 5.92 to 5.09, while matching PixelREPA's reported FID of 4.00 at epoch 200. These results suggest that PatchAFR complements PixelREPA by further improving training efficiency without sacrificing its later-stage performance.
EpochPixelREPA (Paper)PixelREPA (Reproduced)PatchAFR + PixelREPA
1005.95.925.09
2004.0N/A4.00
  • PatchAFR is sensitive to budget prior. The JiT-B/16 results are analyzed in Appendix Fig. 7 and the section "Budget Sensitivity on B/16" of our manuscript. Our experiments show that the effectiveness of the router depends on the frequency budget prior. A suitable prior can improve routing performance, while an unsuitable one may lead to performance degradation.

  • Patch 32 settings are important. Our work primarily focuses on improving the training efficiency of PDMs with 32×32 patches. This configuration has also been explored in many recent works, including miniT2I, JiT, DiP, WiT, DuSPiT, EPG, PixelU, and pMF. More closely related, PixelDiT also investigates patch32 configurations at 256×256 resolution. Given the training challenges of lightweight PDMs with large patches, improving their performance and training efficiency remains an important research direction.

W2. Evaluation across different backbones

We appreciate your idea to generalize it to other backbones. Following some prior work like spectral Forcing, pixelREPA, we only conducted experiments on JiT. We’re also considering it as future work.

Q1. Contribution of Band 3.

Thank you for the insightful question. We conducted an additional ablation study to examine the contribution of Band 3. Specifically, we applied a coefficient {1,0.5,0} to the gate of Band 3, where α=1\alpha=1 corresponds to the original PatchAFR, α=0.5\alpha=0.5 reduces its contribution, and α=0\alpha=0 removes its contribution.

The results on PatchAFR-B/32 show that reducing the contribution of Band 3 consistently increases FID and decreases IS. The visual comparisons further reveal a noticeable loss of fine-grained details when Band 3 is suppressed. These results demonstrate the importance of Band 3 in preserving high-frequency details and generation quality. We have added the discussion "Band 3 Contributes to High Frequency Details" to Appendix A, along with visual comparisons in Figure 11.

SettingFID ↓
Baseline (α=1)11.38
α=0.511.59
α=011.83

Q2. Longer training and different model sizes.

Your insight perfectly aligns our experimental settings and results presented in Table 2 and Figure 4 in the manuscript. Our work focuses on improving training convergence, so we primarily evaluate performance at early training checkpoints. We also investigate longer training schedules and different model sizes in the manuscript.

  • Longer training. Figure 3 presents training results up to 600 epochs, where a variant of PatchAFR consistently outperforms JiT and Spectral Forcing from the middle of training onward. As shown in Figure 4, our final design further improves early-stage convergence while achieving comparable performance to this variant at epochs 100–200.

  • Model sizes. We evaluate PatchAFR on both JiT-B/32 and JiT-L/32. As shown in Table 2, PatchAFR consistently improves FID on JiT-L/32 across all reported checkpoints. We also conduct several ablation studies on L/32 in Tables 3, 4, and 6, further demonstrating that the effectiveness of PatchAFR extends beyond the B-sized model.

Thank you again for your thoughtful questions and suggestions. In particular, W1, W2, and Q1 have helped us further refine our manuscript and strengthen our experimental analysis. We really appreciate your feedback.

Thanks for the additional experiments and clarifications. The Band 3 ablation and the new B/16 + PixelREPA results are helpful. However, following concerns still affect my rating.

  1. Comparison with recent methods. The experiments are still limited to JiT, without comparisons with recent SOTA methods such as PixelDiT [1] and EG-FM [2], which achieve FID scores of 1.54 and 1.45, respectively.

  2. Resolution and task generalization. All experiments are conducted on ImageNet at 256×256. There are no results at 512×512 or on text-to-image generation, so it is unclear how well the method generalizes to more challenging settings.

  3. Scaling to larger models. The largest model evaluated is JiT-L/32 with only 462M paramenters. More importantly, the FID improvement at 200 epochs drops from 9.1% on B/32 to only 1.5% on L/32. This makes me wonder whether the improvement would become even smaller on XL or H models.

  4. Longer training. From Figure 3, at 600 epochs, PatchAFR achieves an FID of roughly 12.2, compared to 13.0 for JiT, which is only about a 6.2% improvement. Also, this result is based on the earlier variant rather than the final design. Would the improvement still hold with longer training, especially for larger models and higher resolutions?

I appreciate the authors' efforts in addressing my questions, but I think these concerns remain important. Therefore, I will keep my rating.

[1] Yu et al., 2026. PixelDiT: Pixel Diffusion Transformers for Image Generation. CVPR 2026. https://arxiv.org/abs/2511.20645

[2] Tong et al., 2026. Energy-Guided Flow Matching. arXiv:2608.05811. https://arxiv.org/abs/2608.05811

Reviewer_2vr7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper builds on Spectral Forcing and further trains a lightweight router to weight different frequency components of the input during denoising, so as to achieve spatially adaptive frequency-domain information processing. Experimental results demonstrate that it achieves certain improvements in training efficiency and final performance compared with JiT and Spectral Forcing.

Critical strengths

In fact, since the information distribution of natural images is spatially uneven, processing the information of different patches at different frequencies during denoising is an intuitive and effective idea.

Critical weaknesses
  1. The improvement is minor. Although the idea is intuitive, the practical performance is mediocre. In Tab1, without cfg, it achieves only a 1.1 FID improvement over JiT and a 0.6 FID improvement over the improved Spectral Forcing. In Figure3, the training curves largely overlap with no obvious distinction.

  2. The experiments and conclusions of the paper lack strong validation. While we do not oppose the use of LLMs for paper writing and refinement, and the paper explicitly states that LLMs were used for processing, it is clear that the authors cannot take responsibility for the final results as they claim. For instance, fig1 of the paper is evidently AI-generated and has not been finalized by the authors, making it unsuitable for human reading (e.g., the text is largely illegible). This raises serious concerns that the authors merely generated the paper automatically with AI without any revision. Furthermore, we doubt whether the authors can truly vouch for the validity of the experiments and conclusions throughout the paper.

Questions for authors

None

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2vr7’s Critical Weakness 2 and Ethics Concerns

We chose to submit our work to ICLR 2027 through OpenReview because we believe in its open, fair, transparent, and traceable review process. Whether our paper is accepted or rejected, we always respect academic reviews based on facts and evidence. However, we are deeply concerned and disappointed by the serious accusations made by Reviewer 2vr7 without clear evidence. We would like to formally respond to these accusations.

AI Use and Author Responsibility

We have clearly disclosed how generative AI was used in our paper. All AI use strictly follows our statement, and there was no undisclosed use or use beyond what we reported. We have also provided detailed experimental results and the complete experimental code to support verification and reproduction.

Without providing any specific evidence, the reviewer directly claims that the authors cannot take responsibility for their research results. We strongly disagree with this unfair and unsupported accusation.

We take full responsibility for every experiment, every number, and every conclusion in our paper. At the same time, we expect the reviewer to take the same responsibility for their public comments and accusations. We believe this is a basic level of respect for the ICLR 2027 review process.

False Accusation About Figure 1

The reviewer claims that Figure 1 is "evidently AI-generated." This statement is simply false.

Figure 1 was entirely created by the authors manually using Microsoft PowerPoint and exported as a vector PDF. The reviewer can easily zoom in on the PDF to clearly see all the details, including the text. We can also provide the original PowerPoint file as evidence.

What we find unbelievable is that, based on this incorrect claim, the reviewer further suggests that our entire paper was automatically generated by AI without proper human review, and even questions the validity of all our experiments and conclusions.

We strongly reject this kind of behavior, where unsupported guesses are presented as facts and then used to question the authors' academic integrity.

Once again, we want to make it clear that our AI use was strictly limited to what we disclosed in the paper. We have already reported these unsupported accusations and the Ethics Review flag to the PC, SAC, and AC, and requested a fair review of this matter.

Reviewer Responsibility and Confidentiality

We noticed that the reviewer repeatedly uses "we" instead of "I" when expressing their opinions and doubts. This gives us reason to question whether other people may have been involved in preparing this review without permission, and whether our unpublished manuscript may have been shared or discussed with others.

Given ICLR's strict rules on independent reviewing and confidentiality, we believe this concern deserves attention, and we ask the reviewer to provide a clear explanation. If our manuscript was shared without permission or other people were involved in the review without authorization, we request that the PC, SAC, and AC formally investigate this matter.

Our Final Position

We fully accept fair, professional, and evidence-based academic criticism. We are also very grateful to the other two reviewers for their constructive feedback, which has genuinely helped us improve our work.

We also respect Reviewer 2vr7's different opinions on the experimental performance.

However, we strongly reject baseless accusations against our academic integrity without any supporting evidence. We also cannot accept an Ethics Review flag based on clearly incorrect factual claims.

We strongly protest these accusations and formally request that the PC, SAC, and AC review this matter to ensure that academic reviews are based on facts and evidence, rather than personal guesses.

AuthorsOpenReview ↗

Response to Reviewer 2vr7's Critical Weakness 1

Thank you for your comments on W1. We have also discussed these results in our responses to Reviewers WQB8 and 5DF9.

We would like to clarify that the improvement should be evaluated not only by the absolute FID difference, but also by the relative percentage of improvement. As shown in Table 2, PatchAFR reduces FID by 8.4%, 7.7%, and 9.1% at epochs 60, 120, and 200, respectively, compared with JiT-B/32.

We believe these results demonstrate meaningful improvements in both generation quality and training efficiency.

Thanks for the response.

  1. The mention of “vector PDF” somewhat addresses our concerns regarding AI generation. That said, we still find it puzzling that the figure contains an excessive amount of small text, as if the authors had not carefully inspected it. This leaves us unconvinced by your statement that “We take full responsibility for every experiment, every number, and every conclusion in our paper.” We wonder whether the authors have read through their manuscript carefully.|

  2. The term “relative percentage of improvement” is not standard in this research field and can lead to misleading evaluations. For instance, an improvement of 1.1 when the FID value is around 11 is negligible, whereas an improvement of 0.2 for an FID value near 2 can be substantial.

Thanks for the comments.

W1: ethics

We strongly disagree with the reviewer's comments regarding our research ethics. We find it difficult to understand how the font size in Figure 1 leads to doubts about whether we have carefully read our manuscript or take responsibility for our experiments, results, and conclusions. A font size issue does not justify questioning the authors' research ethics.

The ICLR 2027 Reviewer Guidelines require reviewers to provide supporting arguments for their assessments. We ask the reviewer to withdraw these unsupported accusations about the authors, which are based solely on a formatting issue in a single figure. We trust that the AC, PCs, and organizing committee will make a fair and impartial judgment on this matter.

W2: FID improvement

First, our work primarily focuses on accelerating training rather than maximizing final performance, as repeatedly emphasized throughout the manuscript. Criticisms centered on relative or absolute FID improvements therefore do not undermine our core contribution.

Second, reducing FID from 12.52 to 11.38 is not a marginal improvement. Our appendix reports the variability of FID measurements in our experiments, showing that the absolute improvement of 1.14 substantially exceeds the observed standard deviation of approximately 0.05. Furthermore, improvements of comparable or even smaller magnitudes have been consistently recognized in the community and reported as main results in papers accepted at top-tier conferences. Examples include:

13.11 → 11.67: DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling.

9.7 → 9.4 and 9.6 → 8.7: Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 Spotlight).

9.6 → 9.33: REPA Works Until It Doesn't: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025).

13.19 → 11.27: Rectified Diffusion Guidance for Conditional Generation (CVPR 2025).

These examples demonstrate that improvements of this magnitude are well within the range of results considered meaningful by the diffusion modeling community.

Moreover, numerous prior works have reported relative FID improvements in comparable ranges, including Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025), Diffusion Tuning: Transferring Diffusion Models via Chain of Forgetting (NeurIPS 2024), Rectified Diffusion Guidance for Conditional Generation (CVPR 2025), and REG: Rectified Gradient Guidance for Conditional Diffusion Models (ICML 2025). We find it difficult to understand why the reviewer questions the validity of our improvements solely on this basis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.