Self Gradient Forcing: Native Long Video Extrapolation
SGF trains earlier generated video context to become better memory for later frames.
The paper identifies a “historical context-gradient gap” in Self Forcing: future-frame losses do not teach earlier generated latents how to write more useful key-value cache states. Self Gradient Forcing adds a second pass that reconstructs context gradients in parallel, avoiding full backpropagation through the serial rollout. The authors report stronger long-video extrapolation than Self Forcing, especially for identity, layout, and temporal stability. They say a model trained on 5-second windows can extrapolate to videos lasting several minutes. HF Daily Papers' note
The paper identifies a “historical context-gradient gap” in Self Forcing: future-frame losses do not teach earlier generated latents how to write more useful key-value cache states. Self Gradient Forcing adds a second pass that reconstructs context gradients in parallel, avoiding full backpropagation through the serial rollout. The authors report stronger long-video extrapolation than Self Forcing, especially for identity, layout, and temporal stability. They say a model trained on 5-second windows can extrapolate to videos lasting several minutes. HF Daily Papers' note
score 5