LORA training settings
Do you mind sharing your training settings like you did for Z-image LORAs?
Of course, here it is. Remember that the amount of steps and amount of images are in relation of 100 to 1. I usually do 22-25 images so I go for 2500 steps, if you have more images, you would need more steps.
Ok. So for 20 images you would drop down to 2000 steps or would you still keep it at 2500.
I would probably drop it to 2000
Of course, here it is. Remember that the amount of steps and amount of images are in relation of 100 to 1. I usually do 22-25 images so I go for 2500 steps, if you have more images, you would need more steps.
I noticed the resolution set in your config is at 512x, is that right? Is 512 enough to train a lora for Flux Klein 9b?? I tried training a 768x lora recently and it always CUDA OOM'd.
yes, 512 is fine
do remember that if you have bucketing (ai toolkit has) then you don't need only 512x512, you can have vertical or horizontal but the multiplication should be within 262144 pixels (you could use my dataset cutter as it has the snap to good ratio functionality)