Tag
This paper introduces BiDPO, a framework that enhances text-to-image models for complex compositional prompts through preference-based fine-tuning and region-level guidance, achieving state-of-the-art results on compositional fidelity benchmarks.