acceptodds
Under review as a conference paper at ICLR 2027

Simplifying and Improving On-Policy Distillation for Few-Step Diffusion

Abstract

Few-step distillation reduces the sampling cost of diffusion and flow models, but existing methods often rely on auxiliary score models or specialized architectures. In this work, we show that a simple on-policy regression baseline already produces high-quality and diverse samples in 4 NFE by matching teacher velocities along student rollouts, with the student velocity held constant within each sampling interval. To counter the optimization instability caused by the high-curatured velocity field of diffusion teacher, we introduce , a lightweight residual network used only during training to model local velocity variations separately from the student interval mean prediction. We further use weaker predictions from the same teacher as negative references in a repulsive regularizer, which effectively avoid the bad modes from original policy and improve the quality. The resulting framework, , remains entirely regression-based, requires neither online score estimation nor adversarial training, and leaves the generator architecture and sampling procedure unchanged. We validate SMD by conducting extensive experiments on image generation and editing, 3D generation, video synthesis, and joint audio–video generation, with models of up to 33 billion parameters. Using only 4 NFE, SMD outperforms strong few-step baselines and even surpasses its multi-step teachers on several benchmarks, while naturally extending to cross-architecture and multi-teacher distillation. Our anonymous project page is available at https://anonymous.4open.science/w/SMD-Page-15B8/this link.

Primary area:
generative models
Keywords:
diffusion models, flow models, few-step generation, distillation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_CPe2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies few-step distillation for diffusion and flow models. The student predicts a single velocity for each sampling interval and learns from teacher predictions queried within that interval. The authors add two components to this baseline: a training-only Pivot branch that models within-interval velocity variations, and a repulsive term based on weaker teacher predictions. Only the student is used at inference. Experiments cover image generation, image editing, video, 3D, and joint audio-video generation, with additional studies on cross-model and multi-teacher distillation.

Critical strengths
  • The approach is straightforward and does not require an online student-score model or adversarial training. Pivot is removed at inference and therefore adds no inference-time network evaluations.

  • Showing that simple regression can produce competitive four-step results across several large models has practical value. The FLUX GenEval results and the motion activity of Wan-14B are particularly encouraging.

  • The experiments cover a broad range of tasks. The authors also report limitations and mixed results, including cases where cross-architecture supervision hurts performance and the lack of clear statistical evidence for visual improvements in H3. These disclosures help establish the method's scope.

Critical weaknesses

I have three main concerns.

First, the contribution of the repulsive term is not sufficiently clear. Under the paper's stop-gradient implementation, the loss can be rewritten as weighted regression toward an extrapolated teacher target:

veff=vT−λvref1−λ.v_{\mathrm{eff}} = \frac{v_T - \lambda v_{\mathrm{ref}}}{1 - \lambda}.

Under standard linear CFG without nonlinear rescaling, this can also be interpreted as increasing the effective guidance scale. It is therefore unclear what additional benefit the formulation provides beyond adjusting the teacher target or guidance strength. The connection to existing guidance methods deserves a more explicit discussion.

Second, the experiments do not directly establish improved training stability. Tables 5-8 show better scores for some configurations at particular checkpoints, but do not show whether Pivot reduces fluctuations during training. Figure 3 illustrates fluctuations in the baseline without a corresponding Pivot comparison, and each ablation setting uses only one training run. In addition, the main Wan-14B checkpoint uses zero active Pivot residual, so that result does not provide evidence that Pivot stabilizes training at larger scale.

Finally, some improvements are small, and parts of the comparison need clarification. The Qwen comparison includes PDD-Euler, while the four-NFE Midpoint variant reported in the PDD paper is closer to SMD. The ImgEdit improvement from 4.48 to 4.51 also needs to be considered alongside evaluation variability. SMD has lower image diversity than Pi-Flow and PDD, and does not lead on video semantic scores. These results establish a competitive method, but do not yet demonstrate a consistently better overall trade-off.

Questions for authors
  1. How does the repulsive objective differ from extrapolating the teacher target or increasing guidance? After matching the effective guidance strength, is there still an additional benefit?

  2. Could the authors report the baseline, Pivot-only, repulsion-only, and full method under one matched training recipe? For the stability claim, learning curves with and without Pivot, together with independent training seeds on a representative model such as Wan-1.3B, would be helpful.

  3. Why were the reported PDD configurations selected rather than the stronger Midpoint variants? Also, were SMD checkpoints selected using an independent validation set or the same benchmarks used for final reporting?

  4. For small differences such as those in image editing, could the authors provide paired confidence intervals or repeated evaluations to establish whether the improvements are reliable?

3 comments

We thank Reviewer CPe2 for the thorough mathematical and empirical critique, and for recognizing the practical value of our simple on-policy regression framework across diverse tasks. We have updated the manuscript with the requested theoretical clarifications, matched four-way ablations, training-progression comparisons, and multi-run evaluation statistics.

1. Repulsion and CFG

The extrapolated-target formulation holds under squared Euclidean distances, where standard linear CFG with target scale wT=5.0w_T=5.0, reference scale wref=0w_{\mathrm{ref}}=0, and λ=0.1\lambda=0.1 reduces to an effective guidance scale of weff=(wT−λwref)/(1−λ)≈5.56w_{\mathrm{eff}} = (w_T - \lambda w_{\mathrm{ref}})/(1 - \lambda) \approx 5.56. However, SMD optimizes the unsquared L2L_2 distance (Appendix C.2, Eq. 28). For a predicted velocity vv with nonzero residuals, the gradient with respect to vv is

∇vℓ=v−vT∣v−vT∣2−λv−v~T∣v−v~T∣2.\displaystyle \nabla_v \ell = \frac{v-v_T}{\\|v-v_T\\|_2} - \lambda \frac{v-\tilde{v}_T}{\\|v-\tilde{v}_T\\|_2}.

Because the positive and negative residuals are normalized separately by their respective norms ∥v−vT∥2\Vert{}v - v_T\Vert{}_2 and ∥v−v~T∥2\Vert{}v - \tilde{v}_T\Vert{}_2, their relative weighting changes dynamically with the current student prediction rather than reducing to a fixed linear extrapolation of the teacher target. Among student predictions with identical teacher-matching error, those lying closer to the weak reference v~T\tilde{v}_T incur a steeper repulsive gradient, introducing a directional preference motivated by weak-model guidance.

To test whether repulsion provides benefits beyond matching this effective guidance strength, we added a controlled comparison on Wan2.1-1.3B at 4 NFE (Appendix I.3, Table 30 and Figure 30). Holding Pivot and all other training settings fixed, we compare CFG 5.0 with repulsion (λ=0.1\lambda=0.1, wref=0w_{\mathrm{ref}}=0) against CFG 5.6 without repulsion. We also evaluate extracted video frame grids using GPT-5.6 Sol for joint Realism and Saturation (R&S) on a 0–5 scale. Repulsion at CFG 5.0 outperforms CFG 5.6 across VBench Quality, Semantic, and Total scores while achieving a substantially higher R&S score of 3.182 compared to 2.853, avoiding the color over-saturation caused by simply raising teacher CFG.

SettingCFGR&S (0–5) ↑
Repulsion5.03.182
No repulsion, stronger CFG5.62.853

2. Component Ablations and Stability

We have added matched four-way ablations covering the baseline, Pivot-only, repulsion-only, and full SMD across Qwen-Image, Wan2.1-1.3B, and Wan2.1-14B at 4 NFE (Appendix I.1, Tables 28–29). Without repulsion, enabling Pivot improves VBench Total from 83.04 to 83.78 on Wan2.1-1.3B and from 83.48 to 84.19 on Wan2.1-14B. When combined with repulsion, adding Pivot raises Qwen-Image OneIG Overall from 0.534 to 0.540, with Text accuracy increasing from 0.921 to 0.944, and improves Wan2.1-1.3B VBench Total from 84.11 to 84.31. On Wan2.1-14B, repulsion alone already reaches a VBench Total of 84.68; adding Pivot further increases Semantic alignment from 80.52 to 80.80 while trading off a slight shift in Quality from 85.72 to 85.63, resulting in a nearly identical Total of 84.66.

Qwen-Image

PivotRepulsionOverallAlignmentTextDiversityStyleReasoning
OffOff0.5200.8650.8680.1600.4150.291
OnOff0.5200.8680.8810.1430.4110.298
OffOn0.5340.8770.9210.1510.4260.295
OnOn0.5400.8830.9440.1440.4290.300

Wan2.1-1.3B and Wan2.1-14B

PivotRepulsion1.3B Quality1.3B Semantic1.3B Total14B Quality14B Semantic14B Total
OffOff84.5577.0083.0484.6079.0083.48
OnOff85.1578.3183.7885.2380.0284.19
OffOn85.4678.7084.1185.7280.5284.68
OnOn85.6478.9784.3185.6380.8084.66

To directly examine training stability, we added side-by-side checkpoint trajectories with and without Pivot on Qwen-Image and Wan2.1-14B under matched prompts, initial noise, and training iterations with repulsion disabled (Appendix I.2, Figures 28–29). On Qwen-Image, the baseline without Pivot suffers from intermediate structural and text regressions between iterations 300 and 500—such as distorted stop-sign lettering and a neon night skyline temporarily collapsing into a daylight scene—whereas enabling Pivot maintains consistent geometry, readable text, and stable scene appearance throughout training. Similarly, on Wan2.1-14B, Pivot preserves more consistent scene composition across intermediate checkpoints.

Continued response to Reviewer CPe2.

3. Comparison Protocol

PDD Comparison

In our main Qwen-Image table, we quoted PDD-Euler because Euler queries the teacher once per local target, matching SMD's single-velocity target construction, whereas PDD-Midpoint refines each target through an additional midpoint query. For Wan2.1-14B, we reported PDD-Midpoint because the PDD paper does not report an Euler variant at that scale.

When compared directly against PDD-Midpoint at 4 NFE, SMD still achieves higher scores on Qwen-Image GenEval, DPG-Bench, and OneIG Overall, and matches PDD-Midpoint on Wan2.1-14B VBench Quality and Semantic scores. On Wan2.1-1.3B, SMD trails PDD-Midpoint on VBench Total at 84.31 versus 84.94, but substantially outperforms PDD-Euler on Semantic alignment at 78.97 versus 72.50 (Table 3). Notably, SMD achieves these results using a global batch size of 16 for both Qwen-Image and Wan2.1, compared to 2,048 and 256 in PDD.

ModelMetricPDD-MidpointSMD
Qwen-ImageOneIG Overall0.5380.540
Qwen-ImageDPG88.6688.7
Qwen-ImageGenEval0.8600.873
Qwen-ImageOneIG Diversity0.1740.144
Wan2.1-14BVBench Quality85.6985.72
Wan2.1-14BVBench Semantic80.7180.52

Checkpoint Selection

Because evaluating a complete benchmark is time-consuming, we evaluate a small subset of benchmark examples to help select checkpoints. We have clarified this procedure in the appendix.

4. Evaluation Reliability

To account for VLM judge variance in image editing, we re-evaluated the same generated outputs on Qwen-Image-Edit-2511 five times and now report the mean and standard deviation in Table 2 and Table 14. Across repeated evaluations, SMD and Lightning tie on ImgEdit Overall at 4.48±0.024.48 \pm 0.02, while SMD achieves clear gains on specific editing operations such as Removal, rising from 4.52±0.074.52 \pm 0.07 to 4.70±0.054.70 \pm 0.05, and Adjustment, rising from 4.40±0.074.40 \pm 0.07 to 4.56±0.084.56 \pm 0.08, alongside consistently higher mean scores across all GEdit-Bench dimensions. These results confirm that simple on-policy regression delivers competitive 4-NFE editing performance without auxiliary score networks or adversarial losses.

MethodImgEdit OverallImgEdit Rem.ImgEdit Adj.GEdit SCGEdit PQGEdit Overall
Lightning4.48 +/- 0.024.52 +/- 0.074.40 +/- 0.078.95 +/- 0.087.73 +/- 0.098.18 +/- 0.09
SMD4.48 +/- 0.024.70 +/- 0.054.56 +/- 0.088.98 +/- 0.057.79 +/- 0.098.21 +/- 0.07

5. Diversity and Semantics

We agree that the diversity trade-offs observed in image generation do not trivially translate to semantic gains in video generation.

Specifically, SMD relies on a simple regression objective and operates under limited computational resources, utilizing LoRA rather than full-parameter fine-tuning and smaller batch sizes compared to methods like PDD and rCM. Nevertheless, SMD remains highly competitive across tasks, demonstrating that simple on-policy distillation can achieve robust performance under lightweight training constraints. We have also updated our anonymous project page with more qualitative comparisons against LightX2V and AnyFlow across text-to-image generation, image editing, and text-to-video generation.

Thanks for clarifying how the repulsive term works. The new CFG comparison and ablations address my main concerns, so I’m raising my score to 3 (Weak Accept).

Reviewer_tgXoOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies few-step distillation of diffusion models. Its central claim is that the simplest on-policy regression baseline is already a strong distiller, requiring no additional training, architectural changes, or other techniques. The authors also address training instability issues and test in a wide range of experiments.

Critical strengths
  1. Breadth is impressive. The claim is supported by an extensive array of tests.
  2. The claim that the simplest on-policy baseline is already strong is a welcome and valuable claim for a field crowded with all sorts of little tricks.
  3. The method can be easily adopted.
  4. The method supports video generation.
  5. Ablations experiments are clear.
Critical weaknesses
  1. The novelty over exhibiting close methods like PDD is incremental.
  2. "Surpasses its multi-step teachers" is over-claimed.For example, OneIG overall does not exceed the teacher. Teacher and student also use very different budgets (50 vs 4 NFE), so "surpass" should be scoped precisely.
  3. It’s not clear how the repulsive regularizer works.
Questions for authors
  1. For the repulsive term: can you rule out that it is merely a diversity/entropy regularizer?
  2. Does SMD work with different NFEs, eg, 2 or 8?
2 comments

We sincerely thank Reviewer tgXo for the encouraging review and for recognizing the value of establishing a strong, simple on-policy baseline alongside our broad experimental validation. We have revised the manuscript to address your questions and clarify the scope of our claims.

1. Differences from PDD

While both PDD and SMD supervise student rollouts using teacher velocities within coarse sampling intervals, SMD departs from PDD through a progressive design that keeps student inference completely unchanged while enriching training supervision.

At the foundational level, our central finding is that direct on-policy velocity regression with the unmodified generator head already serves as a strong few-step distiller, avoiding the architectural complexity of PDD's parallel multi-head decoder.

Building on this minimal baseline, when high trajectory curvature requires finer intra-interval modeling, Pivot isolates those local velocity variations into a training-only residual branch whose integral over each interval is zero, thereby preserving the exact coarse endpoint with zero inference overhead.

Finally, moving beyond pure imitation, we introduce a weak-reference repulsive regularizer that steers the student away from degraded teacher modes during optimization, providing an error-shaping mechanism absent in PDD.

2. Teacher Comparisons

We agree that claiming SMD surpasses its multi-step teachers in the summary sections was overly broad given the 4-NFE versus 50-NFE budget difference and per-metric variations such as OneIG Overall. We have removed this over-claim from the Abstract and Conclusion, and scoped all teacher comparisons strictly to individual benchmarks and inference budgets in Sections 1 and 3.

3. Repulsion and Diversity

The repulsive regularizer acts as a directional error-shaping mechanism rather than a diversity or entropy regularizer. As detailed in Appendix C, contrasting the target prediction against a weaker reference from the same teacher draws on weak-model guidance to steer student updates away from degraded modes when a single coarse transition cannot match all local velocities exactly. In our Qwen-Image ablations with Pivot enabled, adding repulsion improves OneIG Overall from 0.520 to 0.540, Alignment from 0.868 to 0.883, and Text accuracy from 0.881 to 0.944.

Qwen-Image, 4 NFE

PivotRepulsionOverallAlignmentTextDiversityStyleReasoning
OffOff0.5200.8650.8680.1600.4150.291
OnOff0.5200.8680.8810.1430.4110.298
OffOn0.5340.8770.9210.1510.4260.295
OnOn0.5400.8830.9440.1440.4290.300

4. Inference Budgets

SMD is formulation-agnostic to the number of sampling steps, as the same rollout objective applies to any discretized time grid without altering the generator architecture. In addition to our 4-NFE experiments across all tasks, we evaluated MiniMax H3 at both 4 NFE and 8 NFE in Table 4. Increasing the budget from 4 to 8 NFE consistently improves SMD across prompt alignment, structural coherence, temporal coherence, and audio quality, outperforming LightX2V-DMD v1.1 on five of the six reported metrics at 8 NFE. Exploring 2-NFE distillation under the same framework is straightforward and is a natural next step for future work.

MethodNFEPromptStructuralTemporalNaturalnessClarityAudio Quality
LightX2V-DMD v1.143.7973.6533.6443.5561.8293.133
SMD43.6093.6383.7383.6383.4863.514
LightX2V-DMD v1.183.9223.6253.5973.5283.2093.440
SMD83.8633.8633.8343.6253.6193.500

Thanks for the responses. I have no further concerns or questions.

Reviewer_rv6WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The manuscript proposes a novel few-step distillation method, called Simplified Matching Distillation (SMD). At each sampling interval, the student predicts a constant velocity, while Pivot models local velocity variations. Both are jointly optimized through teacher velocity matching.

Critical strengths

The paper proposes a novel few-step distillation method that does not require distribution matching.

The proposed method has been validated for its effectiveness in generative models for tasks such as image generation, video generation, and 3D generation.

Critical weaknesses

Compared to the teacher model, SMD still suffers from issues related to diversity. Furthermore, the training stability of SMD has not been verified.

Questions for authors

Although Pivot is introduced to mitigate training instability, the paper lacks a direct comparison of training stability with and without Pivot.

2 comments

We appreciate the recognition of SMD's simplicity and broad applicability. We have added the requested direct Pivot comparisons.

1. Pivot Stability

To directly evaluate Pivot's impact on optimization dynamics, we added side-by-side training-progress comparisons with and without Pivot on both Qwen-Image and Wan2.1-T2V-14B (Appendix I.2, Figures 28–29). All pairs share identical prompts, initial noise, and training iterations, with repulsion disabled to isolate the effect of Pivot.

Training Trajectory Stabilization: On Qwen-Image, training without Pivot exhibits noticeable intermediate regressions—for instance, stop-sign geometry and lettering distort at iterations 300–500 before recovering later, and the neon night skyline temporarily collapses into a daylight scene at iteration 500. Enabling Pivot suppresses these transient structural and appearance shifts throughout training. Similarly, on Wan2.1-14B, Pivot helps maintain more consistent scene layout and object structure across intermediate checkpoints.

Matched 4-Way Component Ablations: We also report full four-way component ablations across Qwen-Image, Wan2.1-1.3B, and Wan2.1-14B at 4 NFE under matched training recipes in Appendix I.1. Without repulsion, enabling Pivot increases VBench Total from 83.04 to 83.78 on Wan2.1-1.3B and from 83.48 to 84.19 on Wan2.1-14B. When combined with repulsion, adding Pivot further raises Qwen-Image Overall from 0.534 to 0.540, alongside a gain in Text from 0.921 to 0.944, and improves Wan2.1-1.3B Total from 84.11 to 84.31.

Qwen-Image

PivotRepulsionOverallAlignmentTextDiversityStyleReasoning
OffOff0.5200.8650.8680.1600.4150.291
OnOff0.5200.8680.8810.1430.4110.298
OffOn0.5340.8770.9210.1510.4260.295
OnOn0.5400.8830.9440.1440.4290.300

Wan2.1-1.3B and Wan2.1-14B

PivotRepulsion1.3B Quality1.3B Semantic1.3B Total14B Quality14B Semantic14B Total
OffOff84.5577.0083.0484.6079.0083.48
OnOff85.1578.3183.7885.2380.0284.19
OffOn85.4678.7084.1185.7280.5284.68
OnOn85.6478.9784.3185.6380.8084.66

2. Diversity

In terms of diversity, SMD is slightly less diverse than pure trajectory-regression methods such as Pi-Flow and PDD due to the repulsive regularizer, yet it remains substantially more diverse than mode-seeking DMD-based baselines while achieving higher overall quality. On Qwen-Image, pure on-policy velocity matching without repulsion preserves a diversity score of 0.160, which is close to Pi-Flow's 0.180. Enabling repulsion trades a modest variance reduction from 0.160 to 0.144 for large gains in text accuracy from 0.868 to 0.944 and OneIG Overall from 0.520 to 0.540. Across both FLUX.1-dev and Qwen-Image, SMD consistently preserves much higher diversity than DMD-based SenseFlow and Lightning.

FLUX.1-dev

MethodOneIG OverallOneIG Div.
SenseFlow (DMD)0.3800.156
Pi-Flow0.4190.229
Ours0.4200.209

Qwen-Image

MethodOneIG OverallOneIG Div.
Lightning (DMD)0.5270.108
Pi-Flow0.5320.180
PDD0.5350.192
SMD0.5400.144

We have updated our anonymous project page with more qualitative comparisons against methods including LightX2V and AnyFlow across text-to-image generation, image editing, and text-to-video generation.

Reviewer_rv6WOpenReview ↗

Official Comment by Reviewer rv6W

Thank you for your responses. My concerns have been addressed, and I will maintain my positive rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Updates and Clarifications

We sincerely thank all reviewers for their constructive feedback and valuable suggestions. We are particularly encouraged by the reviewers' recognition of the simplicity and practical value of SMD, as well as its broad applicability across different generative models and tasks. In response to the reviewers' comments, we have made several updates to the manuscript, including additional experiments, an expanded discussion, and improvements to the presentation. The main changes are summarized below.

Additional Experiments on Pivot Training Stability

In response to Reviewers rv6W and CPe2, we have conducted additional experiments on Qwen-Image-20B and Wan2.1-14B to investigate the effect of Pivot on training stability. We provide qualitative comparisons of generated samples throughout training, together with quantitative analyses of benchmark results with and without Pivot. These experiments offer a more direct assessment of Pivot's role in training stability across different models.

Connections to Existing Few-Step Distillation Methods

To address the questions raised by Reviewers tgXo and CPe2, we have expanded our discussion of SMD and its connections to existing few-step distillation methods. In particular, we provide a more detailed explanation of the motivation behind the repulsive objective. We further conduct a guidance-matched ablation on Wan2.1-1.3B, comparing our repulsive objective against directly increasing the teacher's CFG scale without repulsion. Together, these additions clarify the role of repulsion and its distinction from conventional guidance-based approaches.

Evaluation Clarifications and Presentation Improvements

In response to Reviewer CPe2, we have clarified several aspects of our evaluation protocol and experimental comparisons. We explain our choice of PDD-Euler, which more closely matches SMD's single-query teacher supervision. We also revisit the image-editing evaluation to better assess the reliability of small performance differences. Finally, we have reorganized the appendix for better readability, redesigned Figure 1 for improved visual clarity, and added prompt alignment to the MiniMax H3 evaluation. We have also updated our anonymous project page with additional qualitative comparisons against methods including LightX2V and AnyFlow across text-to-image generation, image editing, and text-to-video generation, demonstrating the visual quality and broad applicability of SMD.

In the revised manuscript, blue marks rewritten and reorganized material, particularly in the appendix. Red highlights the added discussion, new experiments, and related revisions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.