NeurIPS 2020

Compositional Visual Generation with Energy Based Models


Meta Review

While the notions (energy models, product of experts) and leveraged algorithms (Langevin MCMC, Contrastive Divergence training) are not novel, all reviewers (and the AC) appreciated this work as the first to investigate and demonstrate the potential of EBMs for compositional generation from independently learned probability distribution concepts, successfully on real images. Authors responded well to the reviewers questions and suggestions, providing convincing additional experimental results, incl. higher quality generations standing the comparison with GANs. Please incorporate these in the final version of the paper as well as all the useful clarifications provided in your response to reviewers. With this, the AC recommends the paper be accepted at NeurIPS.