UNIST

AbstractWe introduce UNIST, the first deep neural implicit modelfor general-purpose, unpaired shape-to-shape translation, in both 2D and 3D domains. Our model is built on au-toencoding implicit fields, rather than point clouds whichrepresents the state of the art. Furthermore, our translationnetwork is trained to perform the task over a latent grid rep-resentation which combines the merits of both latent-spaceprocessing and position awareness, to not only enable dras-tic shape transforms but also well preserve spatial featuresand fine local details for natural shape translations. Withthe same network architecture and only dictated by the in-put domain pairs, our model can learn both style-preservingcontent alteration and content-preserving style transfer. Wedemonstrate the generality and quality of the translation re-sults, and compare them to well-known baselines. Code isavailable at https://qiminchen.github.io/unist/.