Prerequisites
To follow this tutorial, it’s recommended to use our fork of SD:Next.
The official SD:Next will be able to do so once the PR is landed.
Prepare
Prepare you images for training and place them in a directory (e.g. automatic/datasets/cutecats).
Train
In terminal window A:
cd automatic
# launch sd:next in the background for use with the train script
./webui.sh
In terminal window B:
cd automatic
# export HIP_VISIBLE_DEVICES if you have multiple available gpus
# `--type lyco` works too
python3 cli/train.py --type lora --name cutecats --tag cutecats --input datasets/cutecats
# training for sd-xl is also supported by adding --sdxl to cli/train.py
# and add --model to specify the sd-xl checkpoint, or you need to select it in the webui
# currently no_half_vae is used to fix the original vae, which will cost more vram and have lower performance
# so it only works on rx 7900 xtx/xt for now, unless you use the fixed vae (see pr for detail)
Now you will see logs about images being processed according to the steps specified in the --process argument.
The WebUI in the background might download BLIP and DeepDanbooru models for later interrogation.
After those you will soon see logs printed from the kohya_ss script, and the output LoRAs will be stored in automatic/models/Lora.
Tweak
There are many arguments to play with (run cli/train.py --help for the list), and even more arguments can be altered in cli/options.py.
I won’t go through many of them which are commonly seen in many other training tutorials, but you can definitely increase the process.target_size in cli/options.py for higher resolution of input images.