Method aligns image and text embeddings in vision-language models

Original: TEVI: Text-Conditioned Editing of Visual Representations via Sparse Autoencoders for Improved Vision-Language Alignment

Source: arxiv.org ↗