AI photo editors botch over 40% of edits. Adobe tests a fix

In short: a request like “sharpen this photo and rotate it slightly right” often ends in a useless image. The authors of the EditAgent paper (according to The Decoder, a team from Adobe Research and UC San Diego) looked at why, and built an agent that does better.

What they tested

• On an expert-annotated benchmark, a general-purpose image editor produced no usable result in over 40% of cases.
• No judge, the model that rates the results, was reliable across all edit types.
• According to The Decoder, retrying with the same editor barely helped, while specialised tools such as sharpening, rotation and zoom solved roughly two thirds of the failures.

What worked

• Each step is routed to a general or a specialised editor, and every candidate gets a verification record. With it, off-the-shelf judge models beat models fine-tuned for the job.
• Several candidates generated in parallel from the original and judged in one comparison: experts preferred this to step-by-step refinement in 78.5% of decided comparisons, using no more than half the judge calls.
• According to The Decoder the whole setup needs about ten times the compute of a single editor, and verified outputs produced about 26,000 examples for training better editors.
• The authors say EditAgent improved two open editors across three benchmarks.

How to read this

This is a research paper, not a Photoshop feature, and the numbers come from the authors themselves. The lesson for creators is still practical: generating several versions at once and picking the best beats insisting on fixing one. The extra cost matters too, so it makes most sense for work where a wrong edit is expensive. I write more about image and video tools in the Image, video, voice hub.

Sources

EditAgent: Rethinking Agentic Image EditingarXiv · Nov 2, 2026 · Primary sourceAdobe is testing how well AI agents can use editing tools like a Photoshop professional wouldThe Decoder · Nov 2, 2026