AMT

AMT — All-Pairs Multi-Field Transforms (CVPR 2023, arXiv 2304.09790) — is a video frame interpolation network from Nankai University designed around one observation: large motion and occlusions are easier if you stop guessing flows from local patches. AMT builds bidirectional correlation volumes over all pixel pairs, uses predicted bilateral flows to retrieve correlations, and derives multiple fine-grained flow fields from one coarse pair for backward warping. That gives it task-oriented flows that the authors say compete with transformer-based interpolators while staying convolutional — and therefore fast and light.

The official implementation is at github.com/MCG-NKU/AMT with a project page and a Hugging Face Space demo, and the authors report state-of-the-art accuracy with high efficiency across standard benchmarks. It is newer and less battle-tested in community tooling than FILM or RIFE, so treat it as a strong research option rather than the default pick.

Author-reported (CVPR 2023): state-of-the-art on standard interpolation benchmarks with high efficiency; favorable accuracy-vs-efficiency trade against transformer models