Best OneTrainer SDXL DreamBooth Training Configurations
I have been experimenting with Nerogar OneTrainer for days right now.
Recently I have obtained decent results. Better results could be obtained but I am doing still research and it is taking huge time and resources.
Your support therefore extremely important for me so thank you so much.
I have attached 3 presets so far. Download and put them into the training_presets folder.
Don't forget to download optimizer_prefs. json and put into training_user_settings folder as well. Otherwise your ADAFACTOR optimizer settings will be incorrect.
The installation of OneTrainer so easy and it has an easy to use GUI.
Hopefully I will prepare a full tutorial for it.
Clone the repo into any folder : https://github.com/Nerogar/OneTrainer
Double click install. bat
Put our presets in attachments into the training_presets folder
After that double click start-ui. bat
Load our preset according to your GPU VRAM
In the general tab set your workspace directory and cache directory
In the model tab set your SDXL model path and output destination
In concepts tab add your training images and set their training captions. I trained with ohwx man
I don't see sampling is useful. It uses poorly performing sampler by default
In backup you can set your parameters as you wish. Like save after every 20 epoch. Saved checkpoints will be inside workspace/save folder
As I said I will make hopefully a big tutorial for OneTrainer
What Each Config Means
13GB_Best_Settings.json : This file uses xFormers so quality will be a little bit worse
14_5GB_Best_Settings.json : This is the best quality. Uses Gradient Checkpointing and do not use xFormers
18GB_Fast_Training.json : This one uses xFormers and do not use Gradient checkpointing so speed will be best but a little bit quality degrade due to xFormers
Gradient Checkpointing do not affect quality but only make training slower yet uses lesser VRAM
I also added a full resolution comparison.jpg between default SDXL preset of OneTrainer vs my 14_5GB preset vs Kohya best config ( https://www.patreon.com/posts/89213064 ) we already use
comparison.jpg
optimizer_prefs.json
compare12.jpg
compare13.jpg
compare11.jpg
13GB_Best_Settings.json
14_5GB_Best_Settings.json
18GB_Fast_Training.json
trainer config files and full comparison files