Trajectory-Guided Frame Interpolation for Anime-Style Toon Effects
TL;DR
A machine learning–based pipeline for controllable generation of anime-style fluid effects that enables artists to generate temporally coherent animations by specifying only the first and last frames of the target sequence and sparse motion trajectories of selected points.
Abstract
Toon effects in anime are traditionally created either manually or via physical simulation. Manual production is time-consuming, requiring artists to iteratively draw similar frames. Physical simulations, although faster, provide limited artistic control and often fail to capture the stylized motion characteristic of anime, which does not necessarily follow physical laws. To address these challenges, we developed a machine learning–based pipeline for controllable generation of anime-style fluid effects. A major challenge in this domain is the lack of suitable training data. To overcome this, we created two datasets: a large realistic fluid effects dataset and a smaller dataset of anime-style effects created by professional animators. We fine-tuned a state-of-the-art diffusion-based video generation model on these datasets. The resulted pipeline enables artists to generate temporally coherent animations by specifying only the first and last frames of the target sequence and sparse motion trajectories of selected points.