Content image Style image Compare VGG or Robust ResNet Indeed, if we are to accept the theory that adversarial robustness is the reason VGG works out of the box with neural style transfer, surely we’d find some indication in existing literature that VGG is naturally more robust than other architectures. Perhaps adversarial robustness just happens to incidentally fix or cover up the true reason non-VGG architectures fail at style transfer (or other similar algorithms In fact, neural style transfer is not the only pretrained classifier-based iterative image optimization technique that magically works better with adversarial robustness.