How was TAESD trained?

#4
by thomaseding - opened

If it's not too much trouble to answer, I'm very curious how your TAESD was trained. For example was it entirely from scratch (and with what dataset) or did you somehow slurp trained information from the original SD VAE?

I added some basic information on GitHub here https://github.com/madebyollin/taesd/issues/11

(Decoder is trained from-scratch on a mix of laion and danbooru data; encoder is distilled from the original VAE encoder)

Thank you for the reply. Much appreciated

thomaseding changed discussion status to closed

Sign up or log in to comment