Farm3D learns an articulated object category entirely from "free" virtual supervision from a 2D diffusion-based image generator. We propose a framework that employs an image generator, such as Stable Diffusion, to produce training data for learning a reconstruction network from the ground up.