The goal of this work is to learn implicit 3D shape representation with 2D supervision (i.e., a collection of images). To that end we introduce the Universal Differentiable Renderer (UDR) a neural network architecture that can provably approximate reflected light from an implicit neura