TL;DR
One-Forcing enables stable 1-step autoregressive video generation by augmenting DMD-based causal distillation with a shared noised-latent adversarial critic, achieving state-of-the-art 1-step VBench performance and efficient framewise generation.
One-Forcing against one-step causal baselines
Method
Adversarial distribution matching for causal generation
One-Forcing keeps the causal autoregressive generator and DMD objective, then turns the trainable fake-score network into a joint score critic and noised-latent discriminator. Real ODE latents and generated latents are noised at sampled timesteps and passed through the same transformer backbone, so adversarial feedback targets the same distribution used by the DMD gradient.
Gallery
Additional One-Forcing samples
Citation
BibTeX
@misc{feng2026oneforcing,
title = {One-Forcing: Towards Stable One-Step Autoregressive Video Generation},
author = {Feng, Jiaqi and Cui, Justin and Ban, Yuanhao and Hsieh, Cho-Jui},
year = {2026},
eprint = {2605.23458},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
doi = {10.48550/arXiv.2605.23458},
url = {https://arxiv.org/abs/2605.23458}
}