Connected Self Forcing: Beyond Local Learning in Video Autoregression
The Connected Self Forcing (CSF) framework restores gradient flow from later video chunks to earlier ones, solving visual inconsistency issues in long video generation. It achieves this by detaching historical caches and introducing Shortcut Gradient Replay to reconnect cross-chunk gradients. This allows each chunk to learn both its own quality and serve as context for subsequent generations, leading to superior long-horizon consistency.