From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliance

Open in new window