EditAgent Enables General AI to Better Evaluate Images

Adobe Research and UC San Diego present EditAgent, an image editing agent that relies on local views and objective measures to evaluate variants in parallel. With these verification signals, non-specialized language models outperform systems trained specifically for the task. The method requires more computation but generates data that subsequently improve faster editors, a key point for Adobe's tools already in beta.

With verification, seven LLMs outperform fine-tuned models

When each candidate is accompanied by a verification record, the seven tested language models surpassed the fine-tuned models in human evaluations. Without this information, none of the seven matched the best specifically trained model. In this protocol, the judging model had to choose the best rendering among several, and cases where professionals deemed it usable were counted, with references like EditScore for comparison. Qwen3.8-27B selected a usable image 81.3% of the time, compared to 74.8% for EditScore-72B. Researchers believe that what the judging model sees may matter more than the training method.

More accurate but ten times more computationally expensive

EditAgent produces more accurate results than previous agents, at the cost of requiring about ten times more computational power. The validated outputs are used to train a corpus, which includes nearly 26,000 training examples. Editors trained on this dataset achieved better results on three benchmarks, while using data generated directly by FLUX led to a decline in performance. Researchers estimate that this trade-off could make the approach cost-effective in the long run.

Parallel variation outperforms iterative refinement

Out of 600 modifications evaluated blindly by five professionals, the versatile editor Flux-2-Klein-9B failed 42.7% of the time. Specialized tools managed to recover about two-thirds of these cases, particularly on cleaning, rotation, or zoom operations, while multiplying attempts with the same editor yielded little benefit. Researchers contrasted an iterative refinement strategy with parallel variant generation: refinement accumulates small errors as it starts from already modified images. Professionals preferred the parallel variants derived from the original in 78.5% of comparisons, and the agent required only about half the evaluation passes in this mode.

What the judge sees: close-ups, change maps, and measurements

EditAgent plans the query in steps and confronts the versatile editor each time with specialized tools, before submitting all variants to a language model in one pass. Each variant is provided with a verification record that includes a close-up of the modified area, a change map, and measurements of brightness and color. In an example of slight brightening, the versatile editor overexposed and warmed the colors, while the specialized tool stayed within specifications. These local views and statistics reveal errors invisible at the scale of the entire image. The agent also checks for consistency across steps and resolves conflicts, such as when the instruction to increase daylight makes a previously frosted window transparent again.

Direct link with Adobe products and evolution of FLUX models

In June, Adobe introduced AI agents in Photoshop, Premiere, and other Creative Cloud applications in public beta: the user expresses their intent, and the agent orchestrates the steps. The ability to choose the right tools and verify its own output directly relates to these products. Additionally, Black Forest Labs replaced the tested versatile model with Flux 3 Image Klein. This new model has 5.6 billion parameters, fewer than Flux 2 Klein, and according to internal tests by Black Forest Labs, it outperforms all other open models as well as the commercial Flux 2 Pro.