Controlling Diversity in Single-shot Motion Synthesis

Eleni Tselepi, Spyridon Thermos, Georgios Albanis, Anargyros Chatzitofis

Keywords: Motion synthesis, 3D animation, diverse, single-shot, GAN

Abstract:

We consider the task of controllable and diverse motion synthesis
from a single sequence as an alternative to the data-dependent
text-to-motion methods, which pose ambiguities in data ownership
and privacy. Recent works in hierarchical single-shot synthesis
have paved the path for unconditional generation and editing tools,
however the methods that focus on 3D animation have failed to
control the diversity of the generated motions. In this paper we
propose the integration of the variational inference in single-shot
GANs, aiming to encode and control the low-frequency generating
factors of the single motion sample. Our experiment showcases
the ability of our VAE-GAN model to control the diversity of its
generations, while preserving their plausibility and quality.

Submission Number: 45