Adobe and UC San Diego test AI image editing agent

A standard language model evaluates variants using quantitative metrics like brightness and color, outperforming fine-tuned evaluation models. In 600 edits, the general editor failed 42.7% of the time; parallel generation with hard data proved superior to iterative refinement. Verified outputs are used to train better editors, improving performance across benchmarks and outperforming commercial agents like Seedream 4.0 and GPT-Image-1.5.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.