Automating steps of the animation production process using AI-based tools would ease the workload of Japanese animators. Although there have been recent advances in the automatic animation of still images, the majority of these models have been trained on human data and thus are tailored to images of humans. In this work, I propose a semi-automatic and scalable assembling pipeline to create a large-scale dataset containing clips of anime characters’ faces. Using this assembling strategy, I create AniVid, a novel anime video dataset consisting of 34,221 video clips. I then use a transfer learning approach to train a first order motion model (FOMM) on a portion of AniVid, which effectively animates still images of anime characters. Extensive experiments and quantitative results show that FOMM trained on AniVid outperforms other trained versions of FOMM when evaluated on my test set of anime videos.