Learned Equivariant Rendering without Transformation Supervision