Generative inbetweening aims to produce plausible intermediate frames between a first and last keyframe, moving beyond pure interpolation. The challenge grows as keyframes become sparser and the motion between them larger, where existing models tend to lose coherence.
The arXiv paper "Anchoring and Rescaling Attention for Semantically Coherent Inbetweening" directly targets this problem. Judging from its title, the authors modify the attention mechanism to anchor and rescale features, presumably to preserve object identity and semantic meaning across wide gaps.
The abstract notes that prior GI models struggle under these conditions, but the full details of the proposed method and results are not included in the available excerpt. No comparison or independent evaluation is reported here, so the claims should be read as the authors' own.