PhotoAgent: Exploratory Visual Aesthetic Planning with Large Vision Models

With the recent fast development of generative models, instruction-based image editing has shown great potential in generating high-quality images. However, the quality of editing highly depends on carefully designed instructions, placing the burden of task decomposition and sequencing entirely on the user. To achieve autonomous image editing, we present PhotoAgent, a system that advances image editing through explicit aesthetic planning. Specifically, PhotoAgent formulates autonomous image editing as a long-horizon decision-making problem. It reasons over user aesthetic intent, plans multi-step editing actions via tree search, and iteratively refines results through closed-loop execution with memory and visual feedback, without requiring step-by-step user prompts. To support reliable evaluation in real-world scenarios, we introduce UGC-Edit, an aesthetic evaluation benchmark consisting of 7,000 photos and a learned aesthetic reward model. We also construct a test set containing 1,017 photos to systematically assess autonomous photo editing performance. Extensive experiments demonstrate that PhotoAgent significantly outperforms existing methods in both instruction faithfulness and visual quality across a diverse range of editing scenarios.

AI photo editing tools have become extremely powerful, but they still rely on users to carefully describe what they want. In practice, getting a high-quality result often means a long loop of trial and error, with repeated prompt refinement and incremental adjustments until the image feels right. This work introduces a different paradigm. The system acts more like a creative agent than a passive editor. It can analyze an image, reason about what could be improved, and generate a plan of edits. It then applies a sequence of changes such as improving lighting, adjusting composition, removing distractions, or refining the overall mood, while checking whether each step improves the result and revising its plan when needed. To support this behavior, the system learns from real photographs paired with human judgments of visual quality, helping it better match human aesthetic preferences. In user studies, its outputs are consistently preferred over existing AI editing tools. More broadly, this work suggests a shift from instruction-following tools to systems that can understand a visual goal, plan improvements, execute edits, and iteratively refine results in a more human-like way.