ATD-12K is a large-scale animation triplet dataset containing 12,000 richly annotated triplets. The training set contains 10,000 animation frame triplets and the test set contains 2,000 triplets. The dataset contains 101 cartoon clips from 30 series of movies produced by diverse producers, with a total duration of more than 25 hours. The collected videos have high resolution of 1920×1080 or 1280×720.

In addition, in order to evaluate the video interpolation method from multiple aspects, the test set contains rich annotations, including difficulty level, regions of interest of motion, and labels of motion categories.