A new arXiv paper, ACEdit, takes on a specific failure mode in image editing: quality drops as edits are applied one after another. The abstract notes that models produce impressive results for individual edits, but their output often deteriorates over multiple editing turns.

The cause, according to the authors, is that each turn introduces a small distribution shift. These shifts accumulate over successive edits, gradually pulling the image away from the desired result.

The paper's title points to asymmetric conditioning as the proposed remedy, but the abstract excerpt does not explain how that conditioning works. As this is a single source, there are no differing findings to compare.