Brain pathologies can vary greatly in size and shape, ranging from few pixels\n(i.e. MS lesions) to large, space-occupying tumors. Recently proposed\nAutoencoder-based methods for unsupervised anomaly segmentation in brain MRI\nhave shown promising performance, but face difficulties in modeling\ndistributions with high fidelity, which is crucial for accurate delineation of\nparticularly small lesions. Here, similar to these previous works, we model the\ndistribution of healthy brain MRI to localize pathologies from erroneous\nreconstructions. However, to achieve improved reconstruction fidelity at higher\nresolutions, we learn to compress and reconstruct different frequency bands of\nhealthy brain MRI using the laplacian pyramid. In a range of experiments\ncomparing our method to different State-of-the-Art approaches on three\ndifferent brain MR datasets with MS lesions and tumors, we show improved\nanomaly segmentation performance and the general capability to obtain much more\ncrisp reconstructions of input data at native resolution. The modeling of the\nlaplacian pyramid further enables the delineation and aggregation of lesions at\nmultiple scales, which allows to effectively cope with different pathologies\nand lesion sizes using a single model.\n