Adobe is testing how well AI agents can use editing tools like a Photoshop professional would

Your AI-assisted image edits are now significantly more accurate and reliable, even for complex tasks.

Adobe researchers, in collaboration with UC San Diego, have developed a new AI agent for image editing called EditAgent. Unlike previous AI agents that retry failed steps with the same tools, EditAgent functions more like a professional Photoshop user by selecting the most appropriate specialized tool for each sub-task. It generates multiple candidate edits in parallel from the original image and uses a standard language model, augmented with verification data (close-ups, change maps, and statistical measurements), to evaluate and select the best result. This approach significantly improves accuracy, outperforming existing evaluation models and even commercial offerings on benchmarks. While EditAgent requires substantially more computing power than single editors, its verified outputs can be used to train faster, more accurate editing models.

The development of EditAgent signifies a shift in AI image editing from brute-force iteration to a more intelligent, tool-aware process, mirroring human professional workflows. This could lead to AI tools that not only automate but also enhance creative control and quality, directly impacting the usability of AI in professional creative software. The reliance on verification data over fine-tuned models for evaluation suggests a future where explainability and data-driven validation become paramount in AI performance.

The increased compute requirement for EditAgent, while substantial, is framed as an investment for generating high-quality training data. This suggests a long-term strategy where expensive initial processing yields more efficient and effective AI models over time, potentially lowering overall operational costs for AI-powered creative services. This approach could accelerate the development of specialized AI editing tools that are both powerful and accessible.