New Framework Optimizes Text-to-Image Prompts with Type-Aware Repairs

Text-to-image (T2I) models often struggle to accurately interpret complex prompts, leading to issues like incorrect object counts or swapped attributes. Current prompt optimization methods tend to apply a generic rewrite, which isn't always effective for diverse failure types. Researchers have introduced TARA (Type-Aware Repair Allocation), a novel framework that addresses this by first diagnosing the specific semantic failure within a prompt. It then routes each identified issue to a specialized repair operator designed for that particular failure type, before compiling these local constraints into an optimized prompt. TARA has demonstrated superior semantic accuracy on established benchmarks like DSG and TIFA, outperforming existing methods. It also maintains high image quality and operates efficiently, making it a promising approach for enhancing the reliability of text-to-image generation.