The current demo uses CLIP model from openai, which is likely what you are looking for