Building an agentic image generator that improves itself

An agentic image generator was developed to create tailored ad inspirations, using OpenAI's Image API and LLM evaluators. Two approaches were tested: LLM-as-a-Judge and a bounding box method. The LLM-as-a-Judge approach proved more effective, especially when separating text clarity from composition improvements. However, LLMs struggled with precise pixel-level corrections. The project demonstrated LLMs' ability to identify semantic-level image defects but difficulty in translating insights into accurate spatial edits. This system represents progress towards more sophisticated AI-driven image generation with guided improvements.