acceptodds
Withdrawn from ICLR 2027

On-Policy Distillation for Few-Step Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-step diffusion models promise interactive generation, but compressing a long teacher trajectory into a few student transitions often degrades fidelity. Existing approaches expose a trade-off. Trajectory and consistency distillation retain paired teacher targets but commonly query states that the few-step student does not visit, while distribution matching follows student-generated samples but replaces those targets with indirect distributional feedback that can favor modes already represented by the student. We present OPTD, which resolves this trade-off by rolling the student to an inference state and matching one coarse transition to the endpoint obtained by integrating a frozen teacher through the same interval. The base method therefore combines on-policy states with dense paired supervision using one trainable network. We add a boundary-matched GAN that compares the student's next-boundary latent with a same-condition real latent at the identical noise level in the frozen teacher's feature space. One endpoint construction supports image, bidirectional video, and causal video generation. Across image and video experiments, OPTD delivers strong and stable few-step generation quality across the evaluated settings.

Primary area:
generative models
Keywords:
Diffusion Models, On-Policy Distillation, Few-Step Generation, Knowledge Distillation, Fast Sampling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zNf8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes on-policy few-step distillation by matching student and teacher endpoints on student-generated rollouts, combined with adversarial and perceptual objectives. The results show competitive performance on image and video generation.

Critical strengths

The proposed method is simple and seems competitive in practice.

Critical weaknesses

My main concern is that the method offers limited methodological novelty and does not demonstrate a clear advantage over prior approaches.

  1. The on-policy idea for few-step distillation is well established. A particularly close precedent is Imagine Flash [1], whose backward distillation also trains the student using teacher-provided regression targets from student-visited states. This work is neither cited nor discussed, and no empirical comparison is provided.

AnyFlow [2] also uses flow-map backward simulation, though with a different supervision objective. DMD2 [3] similarly uses backward simulation with distributional supervision rather than paired regression but similarly uses adversarial objective.

Basically, OPTD appears to combine an established backward-simulation strategy with naive knowledge distillation. I believe that this contribution is largely incremental and needs strong arguments to support it.

  1. Adversarial and perceptual objectives are also established in step distillation. CTM [4] already combines perceptual supervision with adversarial training, while PCM [5] and AYF [6] also incorporate adversarial objectives. There are also many other works that combine these objectives. Thus, the novelty of adding these objectives is limited.

  2. The results are mixed and do not show a consistent advantage over existing methods. It is not entirely clear if the comparison setups are fully controlled and tuned. The image evaluation would benefit from newer preference metrics, e.g., HPSv3 and VisionReward, and human evaluation.


[1] Imagine Flash: Accelerating Emu Diffusion Models with Backward Distillation, 2024.

[2] AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation, 2026.

[3] Improved Distribution Matching Distillation for Fast Image Synthesis, 2024.

[4] Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion, 2023.

[5] Phased Consistency Models, 2024.

[6] Align Your Flow: Scaling Continuous-Time Flow Map Distillation, 2025.

Questions for authors

I highly encourage the authors to justify why the proposed objective combination should be preferred over dozens of possible combinations of various flow map and distributional objectives.

Could the authors provide broader evaluation of most popular flow map and distributional objectives under the similar controlled setup?

Otherwise, it is difficult to see what OPTD contributes beyond combining well-established step-distillation techniques without clear understanding why this combination should be preferred.

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

This paper proposes On-Policy Trajectory Distillation (OPTD) for few-step generative modeling. The student visits an inference-time state and learns to match a frozen teacher’s endpoint over the next sampling interval. Optional boundary-matched adversarial and perceptual losses refine generation quality. Experiments cover image, bidirectional video, and causal video generation, with competitive results from the GAN-refined variants.

Critical strengths
  • The endpoint objective is simple and well motivated. It combines student-visited states with direct paired supervision without requiring an auxiliary score model.
  • The framework applies across image and video settings. The GAN-refined variants achieve promising distributional quality, including image FID of 15.06 and four-step video VBench of 82.64.
Critical weaknesses
  • The central benefit of on-policy endpoint matching is not isolated. The component ablations vary auxiliary losses but do not compare matched on-policy and off-policy queries. In Table 2, base OPTD scores 72.82 versus 81.15 for DMD2, while OPTD+GAN reaches 82.64. Without an off-policy endpoint baseline using the same GAN refinement, it remains unclear how much improvement comes from the proposed core design.
  • The comparison protocol is insufficiently specified. The paper does not identify the FID reference dataset and sample count, or adequately document baseline training budgets and adaptation details. Training compute is also unreported despite full-interval teacher integration and additional discriminator updates. These omissions limit confidence in the comparative results.
  • The causal-video claims have limited empirical support. The comparison with Self-Forcing consists of two qualitative examples, without quantitative results or an aggregate temporal-consistency analysis. These examples do not establish that OPTD more effectively reduces exposure bias or improves long-horizon generation.
Questions for authors
  • Are matched on-policy versus off-policy endpoint ablations available with the GAN configuration and training budget held fixed?
  • Can the authors specify the FID evaluation protocol and baseline training budgets, and provide any existing quantitative causal-video comparison with Self-Forcing?
Reviewer_GvFQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OPTD distills a diffusion teacher by matching each student transition to a teacher endpoint computed from the same student-generated state. Optional adversarial and perceptual losses improve generation quality. Experiments cover images and bidirectional and causal videos.

Critical strengths

The endpoint objective provides direct supervision on states encountered during student inference. The image results are promising: four-step OPTD+GAN achieves FID 15.06, compared with 18.91 for DMD2.

Critical weaknesses

The central contribution is not sufficiently isolated. The paper lacks a matched comparison between on-policy and off-policy query states with the endpoint objective, teacher solver, data, and training budget held fixed. The independent benefit of the proposed coupling therefore remains unclear.

The video evidence is less convincing. Base OPTD scores 72.82 in VBench Total, below DMD2’s 81.15. GAN refinement raises this to 82.64, still below AnyFlow’s 83.82. Causal-video evaluation is primarily qualitative. These results provide limited support for the broader cross-modality claims.

Questions for authors

Can you provide a controlled on-policy versus off-policy query ablation?

Can you provide quantitative causal-video comparisons and clarify baseline training budgets and the image FID evaluation protocol?

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

Related papers

Discussion (0)

Sign in to comment.