This work presents one of the more significant empirical advances in energy-based modeling for image data. The results on up to 128x ImageNet are good and I expect many future works will continue to build on these ideas to sufficiently train energy-based models and leverage their architectures.