Sign In

Attention Lora Training, keep an eye on it

2
Attention Lora Training, keep an eye on it

Attention ONLINE Lora Training, keep an eye on it

my suspicion "set is not that what you get"

i trained on standard FLUX model
with 485 images i set epoches 16 and numrepeats 5 and Train Batch Size 4
so that in sum are 9700 steps
costs ~8200 buzz

after final training (nothing fails) i wonder because the result are not that better than before with ~6000 steps ...

i open and read the metadata from my last trained file:

"img_count\": 485 - thats ok
ss_num_epochs":"16" - thats ok
"ss_num_train_images":"970" - what means that
batch_size_per_device\": 2 - why not 4
"ss_steps":"7760" - why not 9700

in all cases 7760 steps only cost ~6000buzz

so iam not interested in the result iam interested why teh meta-data say some what different that i set in training -> what went wrong ???

civitai - ticket is on the way ;)

2

Comments