Training code
Your tiny series of models are a godsend. I am sure you have had this request multiple times but do you think it might be beneficial to also release training code alongside the models?
I'd be down to collaborate with you if needed :)
Thanks Sayak! A few people have seemed interested in having public TAESD training code at times, but since Serious Battle-Tested VAE training code already exists (from Robin and others) I haven't really prioritized it. My training setup for these models is usually pretty amateurish (just iterating on a checkpoint in a notebook on a single-GPU instance without any ablations, eval, or tuning) so there's no secret Serious Battle-Tested training codebase waiting for me to release 😅
That said, I did finally take some time today to push a cleaned-up version of the basic training notebook + adversarial loss code to GitHub https://github.com/madebyollin/seraena. This code should be capable of training TAESD(XL) from scratch to a reasonable degree of quality.
I haven't yet spent the time/money to determine if a model trained with this notebook would fully match the quality of TAESD/TAESDXL 1.2, but after training for 100k steps on a 10k-image dataset quality looks close enough that it would probably get there.
Even that is helpful! Thanks so much. I will share this a bit more publicly if I have your permission?
Sure, feel free to share to whoever seems interested! I cleaned up the README a bit and added some checkpoints, also posted on Twitter
Thanks a bunch.