3 ms·
huh, thanks, I'll have to look over the website again do they say anything about their training set?
by jazzyjackson 5y ago
huh, thanks, I'll have to look over the website again
do they say anything about their training set?
- visarga 5y agoThey say: > When training the encoder, we sample from the CLIP and DALL-E datasets (approximately 650M images in total) with equal probability. When training the decoder, upsamplers, and prior, we use only the DALL-E dataset (approximately 250M images).