Comparing a query image to some representative of a set of unaligned imagesfrom a class is a cornerstone task for the investigation of ancient ornaments. Whileconvolutional autoencoders provide a level of invariance to translation, they canonly handle a limited range of transformations and often incur blurriness. Wepropose to increase the invariance to linear transformations and standard fluctuationsby using a spatial transformer, then increase reproduction sharpness byusing a fully-connected autoencoder. We evaluate our approach on challengingancient ornament images with synthetic abnormal distortions. This approach significantlyimproves the accuracy of change localization. This information makesit possible to precisely highlight signed changes in image comparators, helpingdomain experts for more efficient analysis of ornaments.