Transflower

Dance requires skillful composition of complex movements that follow rhythmic, tonal and timbral features music. Formally, generating dance conditioned on a piece music can be expressed as problem modelling high-dimensional continuous motion signal, an audio signal. In this work we make two contributions to tackle problem. First, present novel probabilistic autoregressive architecture models the distribution over future poses with normalizing flow previous well context, using multimodal transformer encoder. Second, introduce currently largest 3D dance-motion dataset, obtained variety motion-capture technologies, including both professional casual dancers. Using compare our new model against baselines, via objective metrics user study, show ability probability distribution, being able attend large context are necessary produce interesting, diverse, realistic matches