Traditional 2D cartoons are typically hand-drawn at low frame rates, presenting unique challenges for cartoon frame interpolation due to their non-linear motion and artificial nature. While deep learning has succeeded in real-life video interpolation, applying these techniques to cartoons has been less effective.

To address these challenges, the training objective shifts from L1 loss to a combination of loss functions that better capture the style differences between frames, producing sharper and more perceptually appealing results. Additionally, a Global Motion Aggregation (GMA) module is incorporated to handle occlusion by propagating motion information to occluded areas.

Built upon the AnimeInterp model, our approach is evaluated on the ATD-12K dataset using LPIPS and visual comparisons, achieving significant improvements in perceptual quality.