🎬 A New Approach to Long Video Generation
The Self Gradient Forcing (SGF) training strategy has been developed to eliminate the "historical context gradient disconnection" problem in video diffusion models. The method utilizes a two-pass scheme: a gradient-free rollout and parallel reconstruction of KV-cache gradients.
🌍 SGF enables efficient memory management during long video generation, avoiding catastrophic costs associated with backpropagation through the entire sequence.
👤 This will allow for the creation of neural networks capable of generating videos several minutes long while maintaining character faces and backgrounds with minimal resources.
Source 1: https://zhuang2002.github.io/SelfGradientForcing/ Source 2: https://arxiv.org/abs/2607.20368