Text-Image Alignment

Text-image alignment studies whether models connect words and visual evidence accurately, including objects, attributes, spatial relations, and fine-grained details in settings such as medical imaging. Common approaches learn shared image-text representations with contrastive or local region-to-text objectives, while text-to-image work adjusts prompts, embeddings, noise, or denoising to better satisfy a prompt. Benchmarks increasingly test compositional prompts, hard negative captions, localized errors, and clinical consistency, and evaluators compare captions with visual evidence rather than relying only on a single similarity score. A central challenge is that gains on broad retrieval or image-quality metrics do not guarantee reliable grounding: subtle distinctions, negation, and benign wording changes can still expose errors or make evaluation rankings unstable.