This work tackles the challenge of generating long, high-quality videos with diffusion models, which are usually limited by costly transformers and short-horizon teachers. We propose a simple method that leverages teacher knowledge and self-generated video segments to guide autoregressive students without retraining on long-video datasets. Our approach preserves temporal consistency, avoids error accumulation, and scales video length up to 4 minutes 15 seconds, equivalent to 99.9% of the maximum span supported by our base model's position embedding, significantly outperforming prior methods on fidelity and consistency benchmarks.

Bi-directional diffusion can be seen as a process of gradually restoring a degraded target. We adapt it to autoregressive generation by having a short-horizon teacher refine the student’s outputs and then distilling these correction knowledge back into the student model.

Both Rolling Forcing and LongLive , as well as ours, are able to generate high-quality videos up to multiple minutes long, which marks a significant advance in autoregressive long video generation compared to previous methods. While all methods adopt a windowed distillation strategy, Rolling Forcing introduces progressively varied noise levels across frames combined with attention sink frames, LongLive employs sink frames with KV recaching for prompt switching, our approach relies solely on historical KV cache without sink frames.

Our code will be released soon. Our work can be reproduced based on Self Forcing by following how to reproduce our work.

We sincerely thank the following work for their exceptional effort.

• Self-Forcing: the codebase we built upon.
• Wan: the base model we built upon.
• CausVid: the codebase that shows asymmetric distillation is possible.
• DMD: the key distillation technique used by our method

Please consider citing our work if it's useful. Together we hope to make long video generation better and longer.

@article{cui2025self,
  title={Self-Forcing++: Towards Minute-Scale High-Quality Video Generation},
  author={Cui, Justin and Wu, Jie and Li, Ming and Yang, Tao and Li, Xiaojie and Wang, Rui and Bai, Andrew and Ban, Yuanhao and Hsieh, Cho-Jui},
  journal={arXiv preprint arXiv:2510.02283},
  year={2025}
}