January 25, 2024
Started empirically experimenting for finding the best training hyper parameters for Stable Diffusion 1.5 based models
Using Kohya SS GUI as trainer script.
Rented 8 GPU machine from RunPod and already doing 8 different trainings right now.
If you have any input / feedback / suggestion let me know please.
Comments (4)
m
mypatreonemailacc · January 26, 2024
Can you share the base config you are using? Are you training at 512x512 resolution or higher?
S
SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts · January 26, 2024
I trained both on 768 and 1024. 768 works best and you can generate very good images at 1024. I will post full config. before testing OneTrainer. OneTrainer uses only 10300 MB VRAM on Windows
c
chriseppe · January 25, 2024
Very interesting. What are the main parameters you between the trainings? I.e. which training parameters in your opinion makes the biggest difference in quality?
S
SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts · January 25, 2024
optimizer, optimizer arguments, learning rate. still testing heavily