Decomposing and Interpreting Image Representations via Text in ViTs Beyond CLIP