PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jiaqi Feng, Justin Cui, Yuanhao Ban, Cho-Jui Hsieh

arXiv:2605.23458v1cs.CVcs.AI

TL;DR

One-step causal video generation must reduce multi-step sampling latency without sacrificing visual quality and motion dynamics. One-Forcing augments DMD with a real-data adversarial objective, achieving a VBench score of 83.76 and stable framewise generation at one-third the training cost of chunkwise distillation.

  • Problem

    Causal video generators still require multi-step denoising, leaving latency as a bottleneck for preserving quality and motion in the one-step regime.

  • Method

    One-Forcing combines DMD with an adversarial noised-latent discriminator that reuses the fake-score transformer backbone and grounds its signal in real video data.

  • Results

    83.76 total on VBench establishes state-of-the-art one-step causal video generation, while framewise generation converges stably at one-third the training cost of chunkwise distillation.

  • Takeaways & Limitations

    The results show that stable one-step framewise autoregressive video distillation is achievable with the proposed objective.

  • Takeaways & Limitations

    Vanilla DMD’s local score-matching signal can compound through autoregressive context, producing blur, weak motion, or temporal drift.

Abstract

from arXiv · show

Recent advances have substantially improved real-time interactive video generation in the autoregressive regime. However, most existing few-step autoregressive video generation methods, often distilled from a corresponding many-step teacher, default to a 4-step sampling configuration, which still incurs considerable latency during deployment and suffers from severe quality degradation when the number of sampling steps is further reduced, particularly in the one-step setting. Trajectory-style consistency distillation methods often produce videos with weak dynamics, while DMD-based approaches, such as Self-Forcing, tend to yield blurry frames. To address this challenge, we propose One-Forcing, a simple yet effective approach which augments the DMD objective with an auxiliary GAN loss for high-quality and efficient one-step video generation. 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 and remaining competitive with strong many-step approaches. We further demonstrate that one-step framewise autoregressive generation can be achieved stably with merely one-third of the training cost of the chunkwise model, a setting that prior methods have failed to achieve successfully.

1 Introduction

One-Forcing addresses the latency and quality challenges of one-step causal video generation by combining DMD with an adversarial objective. It achieves stable framewise autoregressive training and state-of-the-art one-step performance on VBench.

  • Bidirectional video models jointly denoise entire sequences, causing costs to grow with video length and hindering real-time or interactive deployment.
  • Autoregressive generators stream frames or temporal blocks, but multi-step denoising per block remains the primary end-to-end latency bottleneck.
  • Video teacher trajectories concentrate curvature near the high-noise endpoint, explaining why trajectory-based distillation degrades sharply when compressed to one step.
  • One-Forcing combines DMD score matching with an adversarial penalty, reusing the fake-score transformer as a noised-latent discriminator grounded in real data.The shared backbone supplies complementary DMD and GAN gradients without additional network overhead.
  • 200 steps enable stable one-step framewise autoregressive training, requiring one-third the training cost of chunkwise distillation while achieving higher quality.
  • 83.76 total is achieved by One-Forcing on VBench, establishing state-of-the-art one-step causal video generation while remaining competitive with strong many-step approaches.

2 Related Works

Related work spans bidirectional and autoregressive video generation, as well as trajectory-based, distribution-matching, and adversarial approaches for reducing diffusion sampling cost. Bidirectional models offer coherence but poor latency, whereas causal models support streaming and modern distillation methods target few-step or one-step generation.

  • Video Generation Paradigms: Bidirectional video diffusion denoises entire clips with full spatiotemporal attention, achieving strong coherence but quadratic computation with sequence length and impractical per-frame latency.These models are effective for offline synthesis but unsuitable for streaming or interactive scenarios.
  • Video Generation Paradigms: Autoregressive video generators factorize clips into sequential conditional blocks, using a KV cache so only the current block is denoised while past blocks remain fixed.This structure naturally supports real-time streaming and interactive generation.
  • Few-Step Generation: Flow matching, rectified flow, and consistency distillation compress diffusion or flow trajectories into few-step or one-step generation, but require trajectories smooth enough for faithful compression.Consistency-style methods have scaled from images to video.
  • Few-Step Generation: DMD instead estimates a reverse-KL gradient from real- and fake-distribution scores, pushing autoregressive video generators toward the data distribution without following a specific teacher path.Video extensions include windowed self-rolled sequences, reward-weighted matching, and diagonal multi-step scheduling.
  • Adversarial Generation: GANs enable single-pass generation through discriminator-guided training, but standalone models have not scaled to broad text-conditioned video distributions.Modern systems therefore use adversarial learning as a post-training or distillation signal.

3 Method

One-Forcing addresses the instability of one-step autoregressive video generation by combining DMD with an adversarial critic, motivated by trajectory curvature and recursive error propagation. It trains a framewise causal generator whose fake-score network also discriminates noised real and generated latents.

  • Motivation: 92.5% of Wan video trajectory curvature mass occurs at t ≥0.9, concentrating the nonlinear region near the highest-noise endpoint.A one-step sampler must approximate this dominant bend in a single jump, unlike a two-step sampler that can place an intermediate anchor afterward.
  • Motivation: One-step autoregressive score errors recursively propagate because each predicted latent block conditions all subsequent blocks through the causal KV cache.This makes the locality of vanilla DMD’s score-difference signal substantially more damaging than in one-step image generation.
  • Joint score and adversarial critic: One-Forcing augments the DMD objective by turning the trainable fake-score network into a joint diffusion critic and noised-latent discriminator.The frozen bidirectional teacher supplies the real score, while the trainable autoregressive model supplies the fake score and denoises generated latents.
  • Training procedure: K = 5 interleaves one fake-score critic update per iteration with one generator update every five iterations on a separately sampled minibatch.The generator uses the normalized fake-minus-real score difference on noised one-step rollouts, while the critic trains the fake score and real/fake classification.
  • Joint score and adversarial critic: The adversarial branch extracts layer-wise critic features with learned register-token queries and classifies noised real versus generated latents using an MLP logit head.The discriminator objective applies softplus penalties to real and fake logits evaluated under the same conditioning setup.

4 Experiments

One-Forcing achieves strong one-step video-generation performance on VBench and in human preference comparisons, surpassing prior one-step causal baselines and many-step methods. Ablations attribute its gains to initialization and adversarial design choices, while forward-KL regularization harms performance and ASD exhibits discriminator collapse.

  • VBench Results: 83.76 total, 85.22 quality, and 77.91 semantic scores make One-Forcing framewise state-of-the-art among one-step causal methods on VBench.With a single NFE, it also surpasses most many-step baselines using 4–25 denoising steps.
  • Human Preference: 88.4% and 92.7% win rates show human preference for One-Forcing over Self Forcing 1-step and ASD 1-step, respectively.The comparisons aggregate decided votes from three annotators across 50 prompts.
  • Initialization Ablation: 83.76 versus 82.36 total and 52.76 versus 23.61 dynamic degree favor causal ODE initialization over causal CD initialization, while CD reaches 80.50 versus 77.91 semantic score.ODE initialization provides richer motion information, whereas CD initialization benefits semantic alignment but suppresses large motions.
  • Forward KL Regularization: 74.83 total, 75.03 quality, and 1.30 dynamic degree result when λfkl=1 is added, substantially degrading the chunkwise baseline’s 81.60 total score.The deterministic squared-error surrogate for forward-KL regularization is poorly aligned with the distributional objectives.
  • Discriminator Effectiveness: One-Forcing’s discriminator maintains a varying logit gap (µ=1.53, σ=1.20), whereas ASD remains near zero (µ=0.001, max <0.006).One-Forcing distinguishes generated latents from real data, while ASD compares two model-generated distributions with minimal difference.

5 Conclusion

One-Forcing augments DMD-based causal video distillation with an adversarial noised-latent branch that reuses the fake-score backbone as a discriminator. Its one-step generator scores 83.76 on VBench, closing most of the gap to 50-step Wan2.1 at 84.26.

  • Method: One-Forcing adds an adversarial noised-latent branch to DMD-based causal video distillation.The method is described as simple and effective.
  • Method: The shared architecture reuses the fake-score backbone as a discriminator, providing real-data-grounded density-ratio feedback without extra parameters.This design supplies feedback from real data while avoiding additional parameters.
  • Results: 83.76 is the resulting one-step generator’s VBench score, closing most of the gap to 50-step Wan2.1 at 84.26.The comparison is reported directly in the conclusion.

6 Limitations and Future Work

One-Forcing requires real data for discriminator training, unlike data-free methods, but such data are available in standard forcing-like distillation settings. Future work will target higher-resolution and longer-duration generation using efficient attention mechanisms.

  • Limitations: One-Forcing requires real data as the discriminator’s real distribution, unlike data-free methods such as Self Forcing and ASD.The distinction concerns the discriminator’s training distribution.
  • Limitations: Training videos or precomputed representations are already available in standard forcing-like distillation settings.This availability reduces the practical burden of One-Forcing’s data requirement.
  • Future Work: Future work will scale One-Forcing to higher-resolution and longer-duration generation by combining it with efficient attention mechanisms.The stated scaling directions cover both spatial resolution and temporal duration.

A Details of Implementations

One-Forcing builds on Causal Forcing and Wan2.1, combining an ODE-initialized causal generator with frozen real-score and trainable fake-score networks. Its implementation uses flow matching, precomputed latent training data, alternating critic-generator updates, and one-step autoregressive inference with a four-step initial warm-up.

  • Model initialization and networks: The framewise model uses an ODE-trained causal initialization, a frozen bidirectional Wan2.1-T2V-14B real-score network, and a Wan2.1-T2V-1.3B fake-score network.The fake-score backbone also serves as the adversarial discriminator after adding register tokens, lightweight attention blocks, and a classification head.
  • Noise schedule and parameterization: The implementation follows Wan2.1 with a flow-matching scheduler and uses shift factor k = 5 for generator rollouts and DMD/GAN critic timestep sampling.During generation, the model predicts vθ and converts it into a clean latent estimate; training uses one denoising timestep per autoregressive block.
  • Data and prompt processing: Training uses precomputed prompt-sample pairs, with adversarial real samples from the dataset and fake samples generated by the current one-step causal generator.The reported framewise model is trained on 21 latent frames with spatial latent size 60 × 104 and 16 latent channels.
  • Training details: Each iteration updates the critic once, while every five iterations additionally updates the generator using the DMD surrogate and adversarial generator loss.The real-score model uses classifier-free guidance scale 5.0 for DMD.
  • Inference details: Inference generates one latent frame per autoregressive block, using one denoising update after a four-step warm-up for the first block to initialize the KV cache.Unless otherwise specified, videos are decoded at 832 × 480 resolution and 16 FPS.

B Trajectory Curvature Analysis Details

The appendix details trajectory-curvature sampling for video and image comparisons and shows that curvature mass concentrates overwhelmingly at high noise, including under a motion-focused metric.

  • Sampling Setup: The video comparison uses 100 50-step Wan2.1-T2V-1.3B trajectories with diverse prompts, shift 8, classifier-free guidance 6, and one deterministic seed per prompt.The image comparison uses eight 256-step trajectories from the official EDM2 ImageNet-512 teacher.
  • Curvature Concentration: 92.49% ± 0.13% of curvature mass lies at t ≥0.9, with a high-noise/mid-noise ratio of 33.1 ± 0.7 across prompts.The 95% bootstrap confidence intervals are [92.24, 92.73]% and [31.8, 34.4], respectively.
  • Motion-Focused Analysis: 88.6% of curvature mass remains at t ≥0.9 under a temporal-difference metric, with a high/mid ratio of 19.3.This metric removes static appearance and emphasizes motion structure.

C Training Loss Curves

Figure 5 compares One-Forcing and ASD training losses over the first 100 steps from ODE-initialized generator checkpoints. One-Forcing stabilizes at a lower DMD-loss level and maintains active generator GAN-loss variation, unlike ASD’s flatlined GAN loss.

  • DMD loss: One-Forcing stabilizes at a lower DMD loss than ASD after similar initial magnitudes.The DMD loss drives the score-matching component.
  • Generator GAN loss: One-Forcing’s generator GAN loss varies actively, while ASD’s GAN loss flatlines.The discriminator provides meaningful gradients to One-Forcing during training.
  • Training loss curves: Figure 5 reports DMD, generator GAN, critic, and discriminator loss curves over the first 100 training steps.The curves compare One-Forcing in blue with ASD in red; both methods start from ODE-initialized generator checkpoints.

D VBench Scores Across All Dimensions

Across all 16 VBench dimensions, One-Forcing outperforms one-step causal baselines on aggregate and several content and dynamics dimensions while maintaining high temporal smoothness. It also exceeds four-step Self Forcing on normalized total and quality scores, with additional dynamic-degree gains.

  • D VBench Scores Across All Dimensions: One-Forcing improves aggregate VBench performance over one-step causal baselines, including Causal-Forcing 1-step, ASD, and Self Forcing DMD 1-step.Figure 6 reports the full 16-dimensional normalized VBench profile for selected Table 1 entries.
  • D VBench Scores Across All Dimensions: One-Forcing shows stronger object, spatial-relation, scene, and dynamic-degree performance than ASD, Causal-Forcing, and one-step Self Forcing.The comparison is made across the selected models’ normalized VBench sub-metric scores.
  • D VBench Scores Across All Dimensions: One-Forcing maintains high temporal smoothness while improving content and dynamics dimensions over one-step causal baselines.The full profile covers all 16 VBench dimensions.
  • D VBench Scores Across All Dimensions: Compared with four-step Self Forcing, One-Forcing achieves a higher normalized VBench total and quality score, with gains in dynamic degree.Figure 6 visualizes higher normalized sub-metric scores as larger radial values.

E Broader Societal Impact

Efficient autoregressive video generation offers benefits for creative, accessible, and interactive applications, but its lower latency also increases misuse risks, making responsible deployment essential.

  • Risks: Low-latency video synthesis can facilitate fabricated media, disinformation, impersonation, and amplification of training-data stereotypes or biases.Lower computational cost reduces a practical barrier to large-scale misuse.
  • Benefits: Efficient autoregressive video generation can support creative production, accessibility tools, rapid prototyping, simulation, and interactive world modeling.The passage frames these as beneficial applications of the technology.
  • Safeguards: Responsible deployment should combine dataset and prompt filtering, provenance tracking, watermarking or content credentials, synthetic-media detection, disclosure, and policy constraints for sensitive uses.These safeguards are presented as practical measures for deployment.
Loading 2605.23458v1…