3 ms·
tldr: bigger text encoder is better. SD will catch up quickly, as conditioning on a new set of precomputed text embeddings is a trivial change
by lucidrains 3y ago
tldr: bigger text encoder is better. SD will catch up quickly, as conditioning on a new set of precomputed text embeddings is a trivial change
- mkaic 3y agoI didn't think about that but you're totally right, assuming they have those embeddings cached it would be super easy to retrain SD using them. 11B parameter count is rather unfortunate though tbh, I've never been the biggest fan of "scale is all you need" even though it seems to ring irritatingly true most of the time.