FiRe: Fine-grained Multimodal Reasoning for Enhanced Image Generation