CNNs perform remarkably well when the training and test distributions are\ni.i.d, but unseen image corruptions can cause a surprisingly large drop in\nperformance. In various real scenarios, unexpected distortions, such as random\nnoise, compression artefacts, or weather distortions are common phenomena.\nImproving performance on corrupted images must not result in degraded i.i.d\nperformance - a challenge faced by many state-of-the-art robust approaches.\nImage corruption types have different characteristics in the frequency spectrum\nand would benefit from a targeted type of data augmentation, which, however, is\noften unknown during training. In this paper, we introduce a mixture of two\nexpert models specializing in high and low-frequency robustness, respectively.\nMoreover, we propose a new regularization scheme that minimizes the total\nvariation (TV) of convolution feature-maps to increase high-frequency\nrobustness. The approach improves on corrupted images without degrading\nin-distribution performance. We demonstrate this on ImageNet-C and also for\nreal-world corruptions on an automotive dataset, both for object classification\nand object detection.\n