acceptodds
Under review as a conference paper at ICLR 2027

Spectral Amplitude Purification in Distribution Matching for Diffusion Distillation

Abstract

Distribution Matching Distillation (DMD) enables high-quality diffusion sampling in only a few steps, but its optimization dynamics remain dominated by coarse, low-frequency signals, delaying the recovery of fine-grained details. We identify a pronounced concentration of spectral amplitudes at low frequencies in the DMD directional error, where dominant low-frequency components overwhelm weaker mid- and high-frequency signals. To address this issue, we propose **S**pectral **A**mplitude **P**urification for Distribution Matching Distillation (SAP-DMD), a plug-and-play approach that adaptively modulates the amplitude spectrum of the DMD directional field. By suppressing the dominant tail of the amplitude spectrum, SAP-DMD reduces low-frequency dominance and promotes more effective recovery of fine structures and textures. Experiments on PixArt-, SD3, and SD3.5 demonstrate that SAP-DMD accelerates training convergence and improves generation quality under both 2-step and 4-step sampling.

Primary area:
generative models
Keywords:
Diffusion Models, distillation, distribution matching, amplitude

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_sHG4OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose SAP-DMA through diagnosing a low-frequency dominance in the DMD method. The proposed method is simple, training-only, and plug-and-play. Extensive experiments demonstrate consistent gains over DMD2 and several competitive few-step baselines on PixArt-α, SD3, and SD3.5 for both 4-step and 2-step regimes, with negligible overhead.

Critical strengths
  • The motivation is clear. The paper is motivated by a clean frequency-domain diagnosis on existing method.

  • The paper introduces a focused, frequency-domain preconditioning of the DMD directional error that is orthogonal to architectural changes and training objectives.

  • Experiments are solid. The authors provide results on three strong backbones (PixArt-α, SD3, SD3.5) and two sampling budgets (2-step, 4-step).

  • The paper is well-written and easy to follow.

Critical weaknesses
  • The authors compare different methods mainly through quatatitative numbers or give image comparison. It would be great if the paper further introduces user study comparison results.

  • While comparisons are broad, some baselines come from re-implementations and others from reported number. it would help to further standardize settings for different methods.

Questions for authors

NA

Reviewer_yQE5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spectral Amplitude Purification for Distribution Matching Distillation (SAP-DMD), a frequency-aware approach to improving few-step diffusion distillation. The authors identify the dominance of low-frequency components in the DMD directional error, which may delay the recovery of fine-grained details during training. To address this issue, SAP-DMD adaptively suppresses dominant spectral amplitudes while preserving phase information, thereby increasing the relative contribution of weaker frequency components. Experiments on PixArt-α, SD3, and SD3.5 under 4-step sampling, together with 2-step experiments on SD3.5, demonstrate improvements in training convergence and generation quality.

Critical strengths
  • The paper provides an interesting frequency-domain perspective on DMD optimization. Analyzing the spectral characteristics of the directional error offers a useful explanation for the delayed recovery of fine-grained details during training.
  • SAP-DMD is simple and introduces negligible computational overhead. The experiments across multiple diffusion backbones, together with the spectral analyses and ablation studies, provide evidence that amplitude modulation can improve few-step distillation.
Critical weaknesses
  • The rationale for suppressing dominant low-frequency components from the beginning of distillation is not sufficiently established. The relative importance of low- and high-frequency corrections may evolve during training, so the benefit of increasing the relative influence of high-frequency signals may depend on the training stage. Although SAP-DMD demonstrates faster fine-detail recovery, it remains unclear whether applying adaptive amplitude suppression from the early stages is preferable to a delayed or progressive suppression schedule. Although the existing analyses provide evidence of robustness, the effects of these schedules on structural fidelity and optimization stability are not directly compared. It also remains unclear whether the reported quality advantage at fixed training budgets remains with extended training as both methods approach convergence, or mainly reflects faster convergence.
  • The methodological novelty over DMD is relatively limited, and the practical benefit of the proposed modification remains unclear. SAP-DMD primarily reweights the spectral amplitude of the existing DMD directional error without substantially changing the underlying framework. Moreover, SAP-DMD preserves the same idealized distribution-matching fixed point as DMD, suggesting that its main contribution lies in modifying optimization dynamics. While faster detail recovery is demonstrated, the improvements over DMD2 in Tables 1 and 4 are relatively modest, making the significance of the final generation-quality gains less convincing.
Questions for authors
  • Why is adaptive amplitude suppression applied from the beginning of distillation rather than introduced later? Could the authors compare against a progressive or delayed suppression schedule and examine its effects on structural fidelity, convergence, and optimization stability? In one representative setting, could the authors also extend the DMD2 and SAP-DMD training curves under matched conditions to assess whether the quality advantage remains as both methods approach convergence?
  • Given that SAP-DMD primarily modifies the DMD directional error while preserving the same idealized distribution-matching fixed point, could the authors clarify its methodological contribution and provide stronger evidence of meaningful final generation-quality improvements over DMD2?
Reviewer_TictOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SAP-DMD to improve few-step diffusion distillation by modifying the directional error in DMD. The authors observe that the spectral amplitudes of the DMD directional error are mostly concentrated in low-frequency components, which they argue could slow down the learning of fine details. Therefore, they propose to adaptively suppress large spectral amplitudes in the Fourier domain while preserving the phase, and use the modified directional error to update the student generator. The method is simple and does not introduce additional inference costs. Experiments on PixArt-a, SD3, and SD3.5 show improvements in generation quality and training convergence.

Critical strengths
  • The proposed method is simple and easy to integrate into existing DMD frameworks. It only requires additional FFT-based operations during training without changing the model architecture or introducing inference overhead. The reported training cost is also nearly identical to DMD2.
  • The results on different diffusion backbones and both 2-step and 4-step generation are promising. Especially for 2-step generation on SD3.5, SAP-DMD achieves a clear improvement over DMD2 on ImageReward and HPSv2.1.
  • The evaluation is quite complete. Besides comparisons with multiple distillation methods, the authors provide ablations on different designs and hyperparameters. Additional analysis on the gradient propagation through the generator Jacobian, high-frequency recovery during training, and robustness across training seeds in the appendix is also appreciated.
Critical weaknesses
  • The current evaluation mainly focuses on preference and text-image alignment metrics. Since the proposed method intentionally suppresses dominant low-frequency signals, I wonder whether it could lead to some negative impact on coarse structures or spatial relationships. Such changes might not be fully reflected by the reported metrics. Additional evaluations on compositional generation, such as T2I-CompBench++, would be helpful to verify that the improved high-frequency recovery does not come at the cost of wrong compositions.
  • Regarding novelty, frequency-aware optimization in diffusion models has been explored in previous works. In particular, FDS [1] and Fourier123 [2] have investigated frequency-aware optimization within score distillation frameworks. Although these methods address different tasks and do not directly perform DMD-based distillation, I feel like a brief discussion on their differences would help better position the contribution of SAP-DMD.
  • In Theorem 1, the authors prove that under idealized situation, the direction field after suppression remains positively aligned with the original one. Based on my understanding, in practice the modified directional error is propagated through the generator Jacobian to update the model parameters. The alignment between the original and modified directions in the output space does not necessarily guarantee the same alignment in the parameter space. Some additional discussion on this would be great.


[1] FDS: Frequency-Aware Denoising Score for Text-Guided Latent Diffusion Image Editing. CVPR'25.
[2] Hybrid Fourier score distillation for efficient one image to 3D object generation.

Questions for authors
  • In Table 4, the authors compare SAP-DMD with spatial-domain clipping and fixed low-frequency suppression, which are valuable ablation studies. I just wonder how much of training stability and convergence improvement actually comes from spectrum rebalancing rather than the reduced overall magnitude of the directional error. Could the authors additionally consider a norm-matched directional-error scaling baseline if possible?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.