Anomaly detection, the task of identifying unusual samples in data, often\nrelies on a large set of training samples. In this work, we consider the\nsetting of few-shot anomaly detection in images, where only a few images are\ngiven at training. We devise a hierarchical generative model that captures the\nmulti-scale patch distribution of each training image. We further enhance the\nrepresentation of our model by using image transformations and optimize\nscale-specific patch-discriminators to distinguish between real and fake\npatches of the image, as well as between different transformations applied to\nthose patches. The anomaly score is obtained by aggregating the patch-based\nvotes of the correct transformation across scales and image regions. We\ndemonstrate the superiority of our method on both the one-shot and few-shot\nsettings, on the datasets of Paris, CIFAR10, MNIST and FashionMNIST as well as\nin the setting of defect detection on MVTec. In all cases, our method\noutperforms the recent baseline methods.\n