3 ms·
Looks like there was a lot of Rothko in the training set
by timmb 6y ago
Looks like there was a lot of Rothko in the training set
- yeldarb 6y agoThe neat thing is: there is no training set! The CLIP model[1] takes arbitrary text prompts. So for these we're just sending eg "An illustration of an upside down dinosaur" and it returns a similarity value between the image's encoding and the text's encoding. To filter out things we don't want (eg text and nudity) we penalize the score based on the similarity for eg "A nsfw illustration". [1] https://openai.com/blog/clip/ https://openai.com/blog/clip/