Recent video generative models have shown strong capability in synthesizing high-quality and temporally coherent videos. However, adapting them for controllable portrait animation is not straightforward, according to the paper's abstract.
The paper introduces CatPortrait, an approach that uses in-context conditioning to tame these video models for portrait animation. The abstract notes that existing methods typically rely on other strategies, but the details are cut off in the available text.
By framing the problem as one of in-context conditioning, CatPortrait aims to leverage the strengths of pretrained video models while gaining control over portrait animation, a task that has remained challenging for prior methods.