Motion inbetweening with score-based diffusion models
||Motion inbetweening with score-based diffusion models|
||Rajmund Nagy <firstname.lastname@example.org>|
||Kungliga Tekniska högskolan|
||2022-08-18 – 2023-03-01|
In recent years, score-based diffusion models have proven to be very strong generative models in several domains, for example text-based image generation. However, they have not been applied to character-animation problems yet. With this project, we aim to develop the first score-based diffusion model that can generate the missing frames of a character animation sequence, a problem known as "inbetweening" in animation. Based on the success of diffusion models in other domains, we expect to significantly push the state-of-the-art forward, in terms of motion quality and temporal coherence, seeing that most current inbetweening models can only generate 1 second of animation. Furthermore, we plan to support fine-grained control over the generated motion by allowing editing and re-generation of the motion sequence. With these benefits, the developed model can rapidly integrated into the workflow of animators in the film and gaming industries.