In this work, we propose a novel method for gen-erating 3D point clouds that leverage properties ofhyper networks. Contrary to the existing methodsthat learn only the representation of a 3D object,our approach simultaneously finds a representa-tion of the object and its 3D surface. The mainidea of our HyperCloud method is to build a hypernetwork that returns weights of a particular neuralnetwork (target network) trained to map pointsfrom a uniform unit ball distribution into a 3Dshape. As a consequence, a particular 3D shapecan be generated using point-by-point samplingfrom the assumed prior distribution and transform-ing sampled points with the target network. Sincethe hyper network is based on an auto-encoderarchitecture trained to reconstruct realistic 3Dshapes, the target network weights can be con-sidered a parametrization of the surface of a 3Dshape, and not a standard representation of pointcloud usually returned by competitive approaches.The proposed architecture allows finding mesh-based representation of 3D objects in a generativemanner while providing point clouds en pair inquality with the state-of-the-art methods.1. IntroductionToday many registration devices, su