Image harmonization is the task of adjusting a pasted object's colour and lighting so it looks native to its new scene. According to the paper, methods are ranked by average error on a single benchmark whose test images are real photographs.
The paper's title, "Only Means Only," signals the central critique: the mean error score aggregates over many test images, and that aggregation can hide where models systematically fail. The abstract indicates the authors are concerned with what the benchmark score actually averages over, rather than taking the number at face value.
Because this is the only source, there is no independent view to compare. This summary is based solely on the abstract, so the specifics of the authors' analysis are not yet available.