tips of the day VAE batch size to
tips of the day VAE batch size to significantly increase latent caching speed in Kohya
5 Replies
Hi, Furkan... this morning I started an SD1.5 train session with the usual 5200 reg images; the training started very fast, so that I tought that reg images were non cached in... and stopped training (in previous training it took nearly 30 minutes to cache in kaggle); does this tip inherit to my case? I read the output to see if reg images were present in the workflow, but I didn't gat the point...
if they are cached
they wont be cached again
Is this something that I should use when I do Dreambooth training on RunPod? To make the caching latents step faster for 5k reg images?
@Dr. Furkan Gözükara
Since there is a
--vae_batch_size flag
And it takes around 20 minutes on a 4090 by defaultyes
increase it to like 4
improves speed a lot
Will do! Thank you!