There isn't just one simple component you can use to do that, but the OpenAI Clip API seems quite accessible. If you don't need to fine-tune CLIP itself, I would recommend calling the CLIP API inside your dataset reader, and making the resulting tensors part of your instances.