Prompt Optimization for T2I Generation

Prompt optimization studies how to turn a user's intent into instructions that image generators follow more accurately, without necessarily changing the generator itself. Common approaches rewrite prompts with language models, inspect generated images for errors using vision-language models, or search over candidate prompts with evolutionary algorithms and reinforcement learning. Methods are evaluated on compositional alignment and instruction-following benchmarks such as GenEval, DrawBench, TIFA, and DSG, as well as specialized goals including style, fairness, and physical consistency. The main tension is improving detail and constraint satisfaction without adding unsupported content, harming image quality, or increasing inference cost.