acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Self-forcing DMD from a Score Sharpness Perspective

Abstract

Self-Forcing with Distribution Matching Distillation (DMD) enables real-time autoregressive video generation, but its outputs can exhibit insufficient motion and degraded visual quality. We revisit these limitations from a score sharpness perspective and use this perspective to motivate changes to the DMD training signal. An illustrative one-dimensional Gaussian mixture connects the noise level to local variations of the score, suggesting a timestep curriculum that gradually shifts training from high-noise to low-noise regimes. We further introduce a timestep offset with antithetic sampling, motivated by the hypothesis that small changes to the teacher’s input timestep may mitigate some prediction errors, and adopt the projection mechanism of Adaptive Projected Guidance (APG) to downweight the parallel guidance component. These three interventions act during training and leave the student architecture and inference procedure unchanged. Experiments on Wan2.1-T2V-1.3B show higher Total, Quality, and Semantic scores than Self-Forcing under the reported evaluation protocol, while preserving real-time throughput. Our method obtains the highest reported Total and Quality scores among the compared methods, with a lower Semantic score than Causal Forcing. Component ablations support the benefits of combining the interventions, including an increase in Dynamic Degree from 23.12 to 56.48 on the ablation prompt suite.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Distribution Matching Distillation, Self Forcing, Autoregressive Video Generation

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_z6f5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper re-examines Self-Forcing for autoregressive video generation through the lens of score sharpness, with the aim of tackling limited motion and degraded visual quality. The authors propose three training-time adjustments: a high-to-low timestep curriculum, antithetic timestep perturbation applied to teacher predictions, and Adaptive Projected Guidance (APG) to temper the guidance signal. Experiments on Wan2.1-T2V-1.3B show that the method improves VBench Total and Quality scores over Self-Forcing and Causal Forcing while keeping inference speed unchanged. Ablation studies further confirm the contribution of all three components, especially in boosting Dynamic Degree.

Critical strengths
  • Good presentation. The paper is well written and easy to follow, and the proposed method is explained clearly and illustrated intuitively.

  • Relatively comprehensive experiments. The paper compares against existing methods and includes detailed ablation studies, which together help demonstrate the effectiveness of the proposed components.

Critical weaknesses
  • Limited novelty and insufficient theoretical justification. The claimed methodological contribution is weakened in two respects. First, the method largely combines existing techniques, namely timestep scheduling, timestep perturbation, and APG. Second, and more importantly, the proposed score sharpness perspective is supported mainly by a simplified 1D Gaussian mixture analysis, and there is no direct evidence linking score sharpness to motion degradation in actual video models.

  • Insufficiently validated performance gains. Although the ablations are extensive, the gains over the strongest baseline, Causal Forcing, remain marginal (+0.12 Total and +0.42 Quality), and Semantic drops by 1.06. Moreover, the lack of multi-seed evaluations, statistical significance testing, and validation across different model architectures makes it difficult to judge how robust and generalizable these gains really are.

Questions for authors
  • Did the authors measure score sharpness in the actual video model, and is there empirical evidence linking it to the observed motion degradation?

  • The teacher receives an input noised at timestep �0��0� but is evaluated at �1��1�. Could the authors clarify why this mismatch helps reduce prediction errors?

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

This paper addresses insufficient motion and visual degradation in Self-Forcing DMD through three training interventions: a timestep curriculum progressing from high to low noise, positive and negative perturbations to the teacher’s input timestep, and an APG-based projection that down-weights parallel guidance. These interventions preserve the student architecture and inference procedure. Experiments with a Wan2.1-based 1.3B student report improvements in visual quality and motion metrics, while a complete component ablation supports the empirical effectiveness of the combined training recipe in the evaluated setting.

Critical strengths
  1. The method offers practical improvements. All three interventions operate during training and can be incorporated into the existing Self-Forcing DMD pipeline. Table 1 reports quality improvements while maintaining 17 FPS throughput and 0.69-second latency. This makes the approach practically relevant for real-time video generation.

  2. The component ablation provides substantial evidence for the combined recipe. Table 2 evaluates all eight component combinations under a stated common training budget and evaluation protocol. The full method improves Quality from 81.22 to 82.90 and Dynamic Degree from 23.12 to 56.48, outperforming every two-component configuration on both metrics.

Critical weaknesses
  1. The score-sharpness perspective does not yet provide a clear, unified explanation of the three interventions. The timestep curriculum has the most direct connection to this perspective: the Gaussian-mixture example illustrates smoother score fields at higher noise levels. However, the analyses of the other two components establish narrower properties. The symmetric timestep expansion cancels first-order temporal variation, but does not establish a reduction in teacher prediction error. The APG analysis conditionally bounds the teacher signal’s magnitude, but does not establish reduced sensitivity of the score to its input. Consequently, the component ablations support empirical usefulness without directly validating the proposed mechanism. The connection between analysis and implementation is also unclear. Equation 9 defines projection in score space, whereas Algorithm 1 also treats the corresponding network output as a denoised prediction. Projection is not invariant between these parameterizations.

  2. The experiments do not establish an advantage over simply reducing CFG strength, and the associated trade-offs are incompletely characterized. Table 5 compares no guidance, default CFG, CFG++, and APG with different projection weights, but does not adequately evaluate intermediate strengths of ordinary CFG. It therefore remains possible that much of the benefit comes from reducing effective guidance strength. Moreover, the guidance ablation does not report semantic alignment, leaving unclear whether improved quality and motion come at the expense of prompt responsiveness. Table 2 also omits Background Consistency and Motion Smoothness, both constituents of standard VBench Quality. Under the standard aggregation, the reported values imply a decrease in the combined contribution of these omitted dimensions, unless a different scoring convention was used. Complete results are needed to determine whether increased motion entails losses in background stability or motion smoothness.

Questions for authors
  1. Beyond the Gaussian-mixture illustration, what evidence from the actual video model supports the score-sharpness interpretation? In particular, how can one determine whether timestep perturbation reduces teacher errors rather than also altering useful prediction structure?

  2. Is APG applied to scores, velocities, or denoised predictions? When querying the teacher at the perturbed timestep t′t', are prediction conversions performed using tt or t′t'? Please reconcile the equations with Algorithm 1.

  3. Could ordinary CFG be evaluated over several guidance strengths under the same training conditions as APG, reporting semantic alignment, quality, and motion together? Does APG retain an advantage at a matched level of prompt alignment?

  4. Please provide all seven Quality submetrics for Table 2, specify their raw values, normalization, and aggregation, and explain whether the omitted background-consistency and motion-smoothness dimensions decrease.

  5. To what extent do the identified issues reflect general limitations of DMD training versus weaknesses specific to the evaluated Self-Forcing configuration? Could the authors test the proposed modifications in another DMD setting, such as non-autoregressive video distillation or a model initialized with Causal Forcing, to clarify their generality?

Reviewer_VS2fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Rethinking Self-Forcing DMD from a Score Sharpness Perspective addresses limited motion and degraded visual quality in Self-Forcing video generation. It proposes three changes to DMD training: a curriculum that shifts from high to low noise levels, a small positive or negative offset to the teacher’s timestep, and projected guidance that reduces part of the teacher’s guidance signal. These changes leave the student architecture and inference procedure unchanged.

Using Wan2.1-T2V-1.3B, the method improves VBench Total and Quality scores over Self-Forcing from 83.25 to 84.16 and 83.90 to 85.01, respectively, at the same reported inference speed. In the ablation study, Dynamic Degree rises from 23.12 to 56.48 when all three changes are combined. The method has the highest reported Total and Quality scores among the compared models, though its Semantic score is lower than Causal Forcing’s.

Critical strengths
  1. Clear, practical method. The three training changes are easy to understand and implement. Each targets a distinct part of the DMD signal, and the paper gives a reasonable motivation for how it could improve motion or visual quality without changing inference.
  2. Systematic ablations. The paper evaluates each change alone, in pairs, and in the full method. It also tests key settings within each component. This makes it easier to see each component’s contribution and supports the claim that combining them gives the strongest overall results under the reported setup.
Critical weaknesses
  1. Limited unifying explanation. The timestep curriculum, teacher timestep offset, and projected guidance act on different parts of DMD training, but the paper does not establish a clear mechanism connecting all three to the proposed score sharpness perspective. The one-dimensional example mainly motivates the curriculum; the other two changes rely on separate hypotheses. The method therefore reads more like a combination of useful remedies than a single coherent solution.
  2. Unclear motion evaluation. The paper reports a Self-Forcing baseline Dynamic Degree of 23.12 on MovieGen prompts, substantially below the 57 reported for Self-Forcing in Causal Forcing[1]. Self Gradient Forcing[2] also reports much higher Self-Forcing Dynamic Degree scores in similar settings. These numbers cannot be compared directly because video length, prompt subsets, checkpoints, and evaluation settings differ. Still, the gap warrants a more precise account of the MovieGen prompt selection and VBench protocol, plus a matched baseline evaluation, before the reported motion gains can be judged confidently.
  3. Limited evidence of generality. All three changes target the DMD training signal rather than a mechanism specific to autoregressive generation. Yet the experiments test only autoregressive text-to-video generation with Wan2.1-T2V-1.3B. Tests in conventional text-to-video and image-to-video distillation would show whether the proposed changes address a broader DMD problem, as the motivation suggests.

[1] Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation

[2] Self Gradient Forcing: Native Long Video Extrapolation

Questions for authors
  1. Can you provide a unifying motivation for the three interventions? In particular, how does the score sharpness perspective explain the timestep offset and projected guidance, and how do the three components work together?
  2. Can you specify the ablation evaluation protocol more precisely, including how many MovieGen prompts were used, how they were selected, and the video and VBench settings? What explains the large gap between your Self-Forcing baseline’s Dynamic Degree and the scores reported in other papers?
  3. Since the interventions modify the DMD training signal, can you evaluate them in other common video distillation settings, such as non-autoregressive text-to-video and image-to-video generation, to test whether the gains generalize?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.