Cartoon Motion Model
In this work, we introduce a framework that enables generation and control of stylized 3D motion by leveraging a pre-trained human motion prior.
Authors
Guilherme G. Haetinger (DisneyResearch|Studios/ETH Zurich)
Dhruv Agrawal (DisneyResearch|Studios/ETH Zurich)
Amit H. Bermano (DisneyResearch|Studios/ETH Zurich)
Robert W. Sumner (DisneyResearch|Studios/ETH Zurich)
Martin Guay (DisneyResearch|Studios)
Jakob Buhmann (DisneyResearch|Studios)
Cartoon Motion Model
Generative 3D motion has shown impressive capabilities for authoring movements, but the integration into animation production is currently hindered by the discrepancy between realistic human motion capture and the exaggerated, non-physical movements of stylized animated feature characters. Nonetheless, hand animated data is sparse, heterogeneous and incomplete, making it a challenging domain for data-hungry generative models. In this work, we introduce a framework that enables generation and control of stylized 3D motion by leveraging a pre-trained human motion prior. Our approach introduces robust data filtering techniques to handle inconsistent and noisy production movements. We train an unconditional ControlNet that eases adaptation by optimizing for maximum alignment between the motion capture and hand-animated domains. The generations of such model are also controllable through a spectral inpainting scheme that exposes parameters for exaggeration, source consistency and multi-character identities. We bridge the gap between physical conformity and artistic stylization and enable a diverse set of applications ranging from retargeting human performances to character-to-character motion transfer