AdaFlow: Efficient Long Video Editing Via Adaptive Attention Slimming and Keyframe Selection

Text-driven long video editing remains challenging due to excessive memory overhead. Recent two-step approaches of keyframe translation and interpolation still face limits on video length. We propose a novel and training-free method, AdaFlow, towards efficient and effective long video editing. By revealing that not all frame tokens are equally important, we introduce Adaptive Attention Slimming to compress the KV sequence, enabling an order-of-magnitude more keyframes. An Adaptive Keyframe Selection scheme further improves quality by choosing representative frames. AdaFlow can edit over 1k frames in a single inference on an A800 GPU, about ten times longer than prior methods like TokenFlow. We also present LongV-EVAL, a new benchmark with high-quality annotations. Project webpage: https://jidantang55.github.io/AdaFlow/.