acceptodds
Under review as a conference paper at ICLR 2027

One-Forcing: Towards Stable One-Step Autoregressive Video Generation

Abstract

Recent advances in autoregressive diffusion-based video generation have substantially improved the quality of real-time video synthesis. However, most existing methods still require multiple denoising steps, while reducing sampling to a single step often leads to severe quality degradation: trajectory-based consistency distillation methods often produce videos with weak dynamics, whereas DMD-based methods, such as Self-Forcing, tend to generate blurry frames. We attribute this limitation to the teacher trajectories exhibiting highly concentrated curvature near the high-noise endpoint, which poses a fundamental geometric challenge for one-step distillation under the consistency distillation framework. To address these limitations, we propose One-Forcing, a simple yet effective approach that augments the DMD objective with an auxiliary GAN loss for high-quality and efficient one-step video generation. We further find that framewise autoregression stabilizes adversarial training, enabling higher-quality generation with substantially fewer training iterations than chunkwise autoregression. Experiments on VBench show that One-Forcing achieves a total score of 83.76, establishing state-of-the-art performance among one-step causal video generation methods while remaining competitive with strong multi-step approaches.

Primary area:
generative models
Keywords:
video generation, world model, generative model, diffusion model

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_o6JxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the problem of one-step autoregressive video generation, where existing distillation methods degrade significantly in quality when sampling is reduced to a single denoising step. The authors make two main observations: (1) trajectory-based consistency distillation fails for video because teacher trajectories exhibit concentrated curvature near the high-noise endpoint (unlike image teachers), and (2) DMD-based methods like Self Forcing produce blurry outputs because low-frequency components dominate DMD updates. To address these issues, they propose One-Forcing, which augments the DMD objective with an adversarial loss using a shared fake-score backbone as discriminator, trained on real video data. They further find that framewise autoregression (one latent frame per block) stabilizes adversarial training compared to chunkwise autoregression. Experiments on VBench show One-Forcing achieves 83.76 total score, outperforming prior one-step causal methods and approaching multi-step quality.

Critical strengths

The trajectory curvature analysis is a novel and actionable diagnostic. The finding that Wan video trajectories concentrate 92.5% of curvature mass at t ≥ 0.9, unlike EDM2 image trajectories, provides a principled explanation for why consistency distillation collapses to one step. This is consequential because it reframes the problem from "distillation is hard" to a specific geometric obstacle, and it motivates the distribution-matching alternative. The rectification experiment (Table 1) provides supporting evidence that trajectory geometry matters.

The framewise vs. chunkwise finding is practically significant and non-obvious. The demonstration that framewise autoregression stabilizes adversarial training (no R1/R2 needed, converges in 200 vs. 750 iterations) and yields better quality (83.76 vs. 81.60) is a useful empirical contribution. The hypothesis about finer temporal credit assignment is plausible and the result is clean. This finding has value beyond this specific method.

Critical weaknesses

Limited evidence for robustness of the main result. The headline VBench score (83.76) is from a single training run with no seed variance reported. Given that the gap to the strongest baseline (ASD at 79.12) is ~4.6 points and the gap to four-step Self Forcing (83.46) is only 0.3 points, knowing the run-to-run variance is important for assessing whether the improvement is meaningful. The extended training result (step 400: 83.93) is reassuring for stability but does not address seed variance.

The one-step curvature claim is not directly tested. The paper argues that high-noise curvature concentration explains why one-step CD fails, but the rectification experiment (Table 1) is conducted at four steps. A direct one-step rectification experiment, or a comparison of one-step CD with rectified vs. original trajectories, would substantially strengthen the causal claim. As presented, the geometric story is compelling but the evidence chain has a gap at the critical operating point.

Questions for authors

Seed variance: Can you report results across at least 2–3 training seeds for the main framewise configuration? Given the small gap to four-step Self Forcing (83.76 vs. 83.46), this would help assess whether the improvement is robust.

One-step rectification: The rectification experiment in Table 1 is at four steps. What happens if you run consistency distillation with rectified trajectories at one step? Does rectification alone recover any of the gap, or does one-step CD still fail? This would directly test the geometric hypothesis at the relevant operating point.

Prompt rewriting consistency: The main results use Qwen-rewritten prompts while some ablations use original prompts. Could you report the main comparison (Table 2) with original prompts as well, or clarify why rewriting was necessary for the main table but not the ablations?

2 comments

Thank you for recognizing the trajectory analysis and the practical value of framewise adversarial training. We address your questions on training-seed variation, one-step rectification and prompt rewriting below.

W1&Q1. Results across independent training seeds.

We have additional framewise runs at 200 iterations, evaluated with one-step generation plus FFE on Qwen prompts:

SeedVBench TotalQualitySemantic
4849183.029184.479377.2286
083.451984.321079.9758

Across the two runs, the mean VBench Total is 83.24, with a standard deviation of 0.30 and a 95% t-confidence interval of [80.55, 85.93]. These repetitions give a clearer picture of variability, showing our method is stable under variant seeds.

W2&Q2. Does rectification help one-step distillation directly?

Thanks for your insightful question!

To obtain one-step students, we adopt progressive distillation. Specifically, starting from the same autoregressive (AR) teacher model released in Causal-Forcing stage 1, we train two four-step students using the original and rectified trajectories, respectively, following the procedure described in the paper. We then progressively distill each four-step student into a two-step student and subsequently into a one-step student, with each stage trained to reproduce the outputs of the preceding model. This controlled setup allows us to isolate and evaluate the impact of trajectory rectification on the final one-step generation performance.

We found that using the rectified trajectory improves the VBench score of the one step student from 55.46 to 57.22. This provides direct evidence that trajectory simplification can help one-step compression, and it is not limited to the original four-step experiment.

Q3. Why use rewritten prompts for the main results and original prompts for the ablation?

Thank you so much for your careful review. Sorry for the confusion. Please see the aligned results below, we will make sure to correct our writings in our future version.

ObjectiveTotalQualitySemantic
DMD-only80.340881.228476.7904
DMD+GAN (ours)83.934385.408278.0388
Gain+3.5936+4.1798+1.2484

The +3.59-point gain confirms an adversarial benefit under the rewritten-prompt protocol as well.

Dear Reviewer,

Given that there is just over a day left in the discussion period, we wanted to reach out to see if our rebuttal above addressed all your concerns or if any others still remained? Please let us know if any further clarification would be helpful. If all your concerns are addressed, we would be really grateful if you could acknowledge that soon.

Thanks, We sincerely appreciate your time reviewing our work and your insightful feedbacks.

Authors

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

One-Forcing combines distribution matching distillation with an adversarial loss on real videos for fast autoregressive video generation. It reuses the fake-score network as a discriminator and generates one latent frame at a time. The paper also studies teacher trajectories to explain the difficulty of one-step consistency distillation. On VBench, the method improves quality over the evaluated one-step causal baselines and shows gains on 20-second videos. The main inference setting uses four denoising steps for the first latent frame and one for subsequent frames, with a reported 1.84x end-to-end speedup over all-four-step sampling.

Critical strengths
  1. The matched adversarial ablation provides clear evidence that real-video supervision helps: adding GAN loss improves VBench from 75.27 to 80.36 while keeping initialization, prompts and FFE fixed (Table 4a).

  2. The reported quality and latency are useful for streaming generation. The main model reaches 83.76 on VBench, improves 20-second rollouts over Self Forcing, and reduces end-to-end time from 7.72s to 4.19s relative to all-four-step sampling, including VAE decoding.

Critical weaknesses
  1. The framewise stability claim needs a stronger controlled comparison. Framewise and chunkwise models use different regularization and training budgets (200 vs. 750 iterations), without matched-compute learning curves or repeated-seed stability results. This leaves a central claimed contribution insufficiently supported.

  2. The trajectory explanation goes beyond the evidence. Equation 3 measures velocity deviation under a chosen time parameterization, and the rectification experiment only evaluates four-step sampling after 300 updates. This does not establish high-noise curvature as the cause of one-step failure.

Questions for authors
  1. Can you compare framewise and chunkwise learning curves at matched wall-clock compute, controlling the regularization settings and repeating training across seeds?

  2. How do the original and rectified CD checkpoints compare under one-step sampling? Does the high-noise concentration remain when the trajectory time parameterization changes?

2 comments

Thanks for your insightful question. We answer your two main concerns below.

W1&Q1. Framewise/chunkwise learning curves, matched regularization, compute and seeds.

We added comparisons with 200 iterations and R1/R2 disabled for both methods, evaluated with matched training and evaluation settings.

SeedFramewiseChunkwiseDifference
4282.149078.9514+3.1976
083.451981.6281+1.8238

We added two paired comparisons with both methods trained for 200 iterations and R1/R2 regularization disabled. Framewise training achieves higher VBench Total in both comparisons, with gains of 3.20 and 1.82 points, averaging 2.51 points. Appendix K, Figure 7 adds the loss and gradient-norm curves; framewise generator gradients show fewer large spikes after the initial phase.

W2&Q2. How do Original and Rectified compare under one-step sampling?

Thanks for your insightful question!

To obtain one-step students, we adopt progressive distillation. Specifically, starting from the same autoregressive (AR) teacher model released in Causal-Forcing stage 1, we train two four-step students using the original and rectified trajectories, respectively, following the procedure described in the paper. We then progressively distill each four-step student into a two-step student and subsequently into a one-step student, with each stage trained to reproduce the outputs of the preceding model. This controlled setup allows us to isolate and evaluate the impact of trajectory rectification on the final one-step generation performance.

We found that using the rectified trajectory improves the VBench score of the one step student from 55.46 to 57.22. This provides direct evidence that trajectory simplification can help one-step compression, and it is not limited to the original four-step experiment.

Q3. Does the high-noise concentration remain after changing the trajectory clock?

To test whether the high-noise concentration depends on the choice of clock, we repeat the analysis using normalized arc length: progress is measured by the cumulative distance traveled in latent space. This coordinate depends on the trajectory’s geometry and is independent of how quickly the path is traversed. We keep the trajectory itself unchanged and use the same physical high-noise region, t≥0.9, under both parameterizations.

On 100 teacher trajectories sampled with 50 solver steps, the high-noise contribution remains nearly unchanged: 92.49% under the original clock versus 90.97% under arc length.

We also repeat the analysis on 20 teacher trajectories sampled with 200 solver steps, obtaining 83.41% under the original clock versus 82.49% under arc length. The agreement between the two parameterizations remains.

These results show that the high-noise contribution persists after removing traversal-speed variation. The one-step rectification experiment provides complementary evidence that simplifying the trajectory can help learning.

Dear Reviewer,

Given that there is just over a day left in the discussion period, we wanted to reach out to see if our rebuttal above addressed all your concerns or if any others still remained? Please let us know if any further clarification would be helpful. If all your concerns are addressed, we would be really grateful if you could acknowledge that soon.

Thanks, We sincerely appreciate your time reviewing our work and your insightful feedbacks.

Authors

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

The paper proposes One-Forcing, which combines distribution matching distillation with adversarial supervision and framewise autoregression for stable one-step video generation. It addresses the blurred details, weak motion, and training instability that arise when autoregressive video diffusion is compressed to a single denoising step. Experiments report a VBench total score of 83.76, outperforming the evaluated one-step baselines and remaining competitive with multi-step models, with additional improvements demonstrated on 20-second videos.

Critical strengths
  1. The quality improvement is supported by a controlled objective ablation. Adding GAN supervision increases VBench Total from 75.27 to 80.36 under matched initialization, framewise training, first-frame enhancement, prompts, and generation seeds. This provides direct evidence for the contribution of adversarial supervision beyond the main baseline comparisons. The ablation uses original prompts and one sample per prompt, with its protocol distinguished from the main evaluation (Tables 2 and 4, Appendix A).

  2. Supplementary evaluations address important practical tradeoffs. Extending training from 200 to 400 iterations preserves generation quality, with Total changing from 83.76 to 83.93. The reported LPIPS and recall measurements provide preliminary evidence against substantial diversity loss under the evaluated protocols. Timing includes VAE decoding and quantifies the first-frame enhancement overhead: 0.06 seconds relative to all-one-step generation, while retaining approximately 1.84× acceleration over all-four-step sampling (Appendices D, H, and J).

Critical weaknesses
  1. The connection between trajectory geometry and one-step distillation remains indirect. The rectification experiment preserves teacher endpoints and controls initialization, sample order, and optimization settings. However, evaluation uses four-step sampling after 300 training iterations. These results support improved learning under that configuration but do not directly establish the effect of rectification on one-step consistency distillation. Consequently, the geometric explanation for one-step difficulty remains less directly validated than the effectiveness of the proposed training objective (Section 3.2, Table 1, Appendix B).

  2. The contribution of framewise autoregression to training stability is not fully isolated. The comparison uses separately stabilized configurations: framewise training without R1/R2 reaches its reported result after 200 iterations, whereas chunkwise training uses these regularizers and reaches its reported result after 750 iterations. This demonstrates an effective framewise configuration, but does not separate autoregressive granularity from regularization and training-budget effects. Matched-budget quality curves or repeated training runs would strengthen the stability and convergence claims (Section 4.3, Appendix A).

Questions for authors
  1. Has trajectory rectification been evaluated directly with one-step consistency distillation? Which conclusions about one-step difficulty are supported by the four-step experiment?

  2. Are matched-budget quality or stability curves available for framewise and chunkwise training? How do their actual training costs compare under the respective regularization settings?

  3. Were the existing one-step baselines in Table 2 reevaluated with identical prompt rewriting, first-frame enhancement, and sampling protocols? If different protocols or previously reported scores were used, could those differences be specified?

2 comments

Thank you for recognizing the controlled adversarial ablation and the practical quality–latency tradeoff. Your questions helped us strengthen the evidence for one-step distillation.

W1&Q1. Has rectification been tested directly at one step? What does the four-step experiment establish?

Thanks for your insightful question!

To obtain one-step students, we adopt progressive distillation. Specifically, starting from the same autoregressive (AR) teacher model released in Causal-Forcing stage 1, we train two four-step students using the original and rectified trajectories, respectively, following the procedure described in the paper. We then progressively distill each four-step student into a two-step student and subsequently into a one-step student, with each stage trained to reproduce the outputs of the preceding model. This controlled setup allows us to isolate and evaluate the impact of trajectory rectification on the final one-step generation performance.

We found that using the rectified trajectory improves the VBench score of the one step student from 55.46 to 57.22. This provides direct evidence that trajectory simplification can help one-step compression, and it is not limited to the original four-step experiment.

W2&Q2. Can framewise and chunkwise training be compared at the same budget and regularization setting? What are their actual costs?

Thanks for pointing it out. We added comparisons with 200 iterations and R1/R2 disabled for both methods, evaluated with matched training and evaluation settings.

SeedFramewiseChunkwiseDifference
4282.149078.9514+3.1976
083.451981.6281+1.8238

We added two paired comparisons with both methods trained for 200 iterations and R1/R2 regularization disabled. Framewise training achieves higher VBench Total in both comparisons, with gains of 3.20 and 1.82 points, averaging 2.51 points. Appendix K, Figure 7 adds the loss and gradient-norm curves; framewise generator gradients show fewer large spikes after the initial phase.

Q3. Were the Table 2 comparisons evaluated consistently? We evaluate all methods in Table 2 using their official code and inference settings, with the same Qwen-rewritten prompts and VBench sampling protocol. FFE is enabled for ASD and our framewise and chunkwise variants, following their official implementations. Self Forcing and Causal-Forcing use their original inference configurations without FFE.

Dear Reviewer,

Given that there is just over a day left in the discussion period, we wanted to reach out to see if our rebuttal above addressed all your concerns or if any others still remained? Please let us know if any further clarification would be helpful. If all your concerns are addressed, we would be really grateful if you could acknowledge that soon.

Thanks, We sincerely appreciate your time reviewing our work and your insightful feedbacks.

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

# General Response to All Reviewers

We thank all reviewers for their constructive feedback and positive recognition of our work. Reviewers highlighted the novel trajectory analysis (o6Jx), non-obvious and practically significant framewise design (o6Jx), convincing adversarial ablations (PkyQ, nPEg), and strong quality–latency tradeoff (PkyQ, nPEg). One-Forcing achieves SOTA one-step causal video generation (83.76 VBench), outperforming prior one-step methods by 4–7 points.

We address the three shared concerns with new experiments:

1. Trajectory analysis (PkyQ, nPEg, o6Jx). We directly tested one-step generation with progressively distilled rectified trajectories, improving VBench from 55.46 to 57.22. The high-noise concentration also persists under arc-length reparameterization.

2. Framewise stability (PkyQ, nPEg, o6Jx). Under matched training settings, framewise outperforms chunkwise by 2.51 VBench points on average across two seeds. Additional runs and gradient curves further support training stability.

3. Adversarial supervision and evaluation (PkyQ, o6Jx). GAN supervision consistently improves VBench by 3.59–5.09 points across prompt protocols. We also clarified that the main baselines share the same evaluation prompts and sampling protocol.

Together, these new results strengthen our main findings and demonstrate the effectiveness, stability, and practical value of One-Forcing.

We really appreciate all the reviewers for reviewing our paper. Please see more details below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.