Anchoring and Rescaling Attention for Semantically Coherent Inbetweening
The authors propose training-free methods to improve generative video inbetweening using two mechanisms: Keyframe-anchored Attention Bias (KAB) and Rescaled Temporal RoPE (ReTRo). KAB uses keyframe cross-attention to guide intermediate frames toward semantic consistency, while ReTRo adjusts temporal attention scaling to stabilize frame fidelity across the sequence. This results in state-of-the-art results on challenging video generation tasks.