2018 Volume 47 Issue 4 Pages 440-446
This paper proposed a motion synthesis and edit for character animations via latent variable space. An existing method using an auto-encoder is unsuited for efficient explorations and manipulations because the latent space has a large number of dimensions. Moreover, its encoder and decoder are composed of a single-layer, which is not suited to synthesize various types or styles of motions. We propose two types of generative neural networks that can map the latent variables so as to fit to a Gaussian distribution and can embed various motions in a lower dimensional latent space. We evaluate the plausibility of motions synthesized with our method, by demonstrating motion transitions and interpolations without preprocessing of time-alignment.