Preprint2026
Self Gradient Forcing: Native Long Video Extrapolation
Junhao Zhuang, Shiyi Zhang, Yuxuan Bian, et al.
Self Gradient Forcing (SGF) restores missing supervision from future losses to earlier generated latents via a two-pass training strategy, enabling native long-video extrapolation beyond the training window.
0Jul 22, 2026Attention MechanismsDiffusion Models
arXiv