We propose a modular inductive bias for disentangled representation learning, which we term a compositional bias, decoupled from both learning objectives and model architectures.