Editor's Choice: Evaluating Abstract Intent in Image Editing through Atomic Entity Analysis

Researchers introduced AbstractEdit, a benchmark for abstract image editing tasks such as mood or style changes.

They also proposed Entity-Rubrics, an entity-level evaluation framework to better judge instruction following and image preservation.

Results show leading image-editing models often struggle with abstract intent, either editing too little or over-editing the image.

The findings highlight a key weakness in multimodal systems and point to the need for stronger language reasoning and better evaluation.