GPT-4: Researchers have introduced LLMScore, a framework that leverages large language models (LLMs) to evaluate text-image alignment in text-to-image synthesis. LLMScore mimics human review by accessing compositionality at various granularities and providing alignment scores with justifications. The system can adapt to different standards, such as overall alignment or error counting, and has demonstrated strong correlation with human judgments across multiple datasets.
Read more at MarkTechPost…