Abstract
Generative modeling of symmetric densities has a range of applications in AIfor science, from drug discovery to physics simulations. The existinggenerative modeling paradigm for invariant densities combines an invariantprior with an equivariant generative process. However, we observe that thistechnique is not necessary and has several drawbacks resulting from thelimitations of equivariant networks. Instead, we propose to model a learnedslice of the density so that only one representative element per orbit islearned. To accomplish this, we learn a group-equivariant canonicalizationnetwork that maps training samples to a canonical pose and train anon-equivariant generative model over these canonicalized samples. We implementthis idea in the context of diffusion models. Our preliminary experimentalresults on molecular modeling are promising, demonstrating improved samplequality and faster inference time.