Interesting, thanks for the link. It seems that CLIP encodings aren't as useful as frozen encoders from the textual domain, which is a little unintuitive imo. Can't keep up with all these advancements!