Dreambooth training taking very long (even for 1000 steps) [4090]
Hi there,
I am running a Dreambooth training that uses sd-scripts under the hood but it seems like it is taking very long on a 4090.
I use the
save_every_n_steps
flag so I can kinda see how long it takes for every 1000 steps. I feel like this is taking very long or am I wrong?
3 Replies

I disabled gradient checkpointing, set the train_batch_size to 4 and the vae_batch_size to 4 aswell to possibly improve the overall training speed and also caching latent step and also enabled xformers
Join @YouTube.com/@SECourses discord and ask there