Compositional T2I Generation

Compositional text-to-image generation studies how models can depict all the parts of a prompt while binding each attribute and relationship to the correct object and region. Methods include attention and layout controls, personalized adapters, structured scene plans, and reasoning or verification loops that detect and repair errors; some improve generation through inference-time guidance, while others train with preference or reinforcement-learning rewards. Benchmarks such as GenEval and T2I-CompBench test object presence, counting, attribute binding, and spatial relations, with newer evaluations probing occlusion, long prompts, and complex layouts. Despite progress, crowded scenes and precise relations remain difficult, and a central tension is improving prompt faithfulness without sacrificing image quality, identity fidelity, or inference efficiency.