New papers building on VLMs: 📎 ProLIP shows that adequately finetuning the last CLIP layer significantly boost few-shot classification. ☕ LatteCLIP proposes an unsupervised prototype-based CLIP finetuning from only synthetic labels.