Adobe Research Publishes Study of an AI That Picks Its Own Editing Tools: Better Quality, 10x the Compute

Researchers at Adobe Research and UC San Diego released a paper on October 29 describing an image editing AI called “EditAgent.” The system breaks an editing request into multiple tasks, then picks a suitable tool for each one.

How it chooses an editing tool

EditAgent generates several candidate edits for a single task and compares them. It applies both a general-purpose image editing model and specialized tools built for that particular kind of task. Earlier research had a smaller pool of tools to draw on, so it often did little more than retry a task that had failed. By widening the range of tool choices, EditAgent creates the opportunity to get a different result.

When selecting among candidates, it does not just look at the finished image. A model that handles both text and images reviews a zoomed-in view of the altered region, a map showing where the changes occurred, and measurements of color and brightness. It also checks whether a later task damaged an edit that had already been completed. The design aims to preserve earlier changes as requests are processed in sequence.

Did multiple candidates help?

In the researchers’ experiments, specialized tools rescued some of the edits that the general-purpose model failed at. Across 600 cases rated by five professional editors, FLUX.2-klein-9B produced no usable result in 256 of them, or 42.7% of the total. The researchers reported that specialized tools produced a usable candidate in 175 of those 256 cases.

Generating candidates all at once also rated better than refining a single result step by step. According to the researchers, in 130 comparisons where a clear preference was expressed, the parallel-candidate approach was chosen 102 times, or 78.5%.

The researchers said they ran three evaluations on top of two publicly available editing models and saw performance improve. From the edits that passed verification, they built roughly 25,700 training examples. An editing model given additional training on that data outperformed a model trained on data generated directly by the base editing model.

Ten times the compute: is it practical?

The price of higher edit quality is computation. According to The Decoder, running EditAgent requires about 10 times the compute of using a single editing model, because it generates and verifies multiple candidates for every task. How that translates into processing time and cost in a real product remains to be seen.

Adobe previously announced a public beta of AI Assistant for Creative Cloud apps including Photoshop in June 2026. What EditAgent aims to change is the part where the AI itself selects the tools and the final result to match a user’s request. Users in South Korea will also want to watch for local availability, how well Korean-language requests are handled, and pricing.