Paper 2510.02283
Self-Forcing++: Towards Minute-Scale High-Quality Video Generation
- Published
- Oct 2025
- Research lab
- Independent
- Citations
- 156
- GitHub
- 268 stars
01 In brief
Summary
Self-Forcing++ is a method for long-horizon video generation that extends autoregressive diffusion models beyond the training horizon of their teacher models.
It addresses quality degradation from error accumulation by generating long self-rollouts (up to 100 seconds), re-injecting noise into these degraded sequences (backward noise initialization), and applying extended distribution matching distillation (DMD) with a short-video teacher on sampled windows.
This approach, combined with a rolling KV cache during both training and inference, avoids recomputing overlapping frames and mitigates over-exposure.
The method also uses GRPO with an optical-flow reward to improve long-term smoothness.
Experiments show it generates videos up to 4 minutes 15 seconds (50x longer than baseline) with superior fidelity and consistency compared to baselines like CausVid and Self-Forcing.
The paper also proposes a new evaluation metric, Visual Stability, using Gemini-2.5-Pro, to address biases in existing benchmarks like VBench.
02 From the paper
Abstract
Diffusion models have revolutionized image and video generation, achieving unprecedented visual quality. However, their reliance on transformer architectures incurs prohibitively high computational costs, particularly when extending generation to long videos. Recent work has explored autoregressive formulations for long video generation, typically by distilling from short-horizon bidirectional teachers. Nevertheless, given that teacher models cannot synthesize long videos, the extrapolation of student models beyond their training horizon often leads to pronounced quality degradation, arising from the compounding of errors within the continuous latent space. In this paper, we propose a simple yet effective approach to mitigate quality degradation in long-horizon video generation without requiring supervision from long-video teachers or retraining on long video datasets. Our approach centers on exploiting the rich knowledge of teacher models to provide guidance for the student model through sampled segments drawn from self-generated long videos. Our method maintains temporal consistency while scaling video length by up to 20x beyond teacher's capability, avoiding common issues such as over-exposure and error-accumulation without recomputing overlapping frames like previous methods. When scaling up the computation, our method shows the capability of generating videos up to 4 minutes and 15 seconds, equivalent to 99.9% of the maximum span supported by our base model's position embedding and more than 50x longer than that of our baseline model. Experiments on standard benchmarks and our proposed improved benchmark demonstrate that our approach substantially outperforms baseline methods in both fidelity and consistency. Our long-horizon videos demo can be found at https://self-forcing-plus-plus.github.io/