Copy of the pin in LORA Training thread from the Furry Discord. No guarantees if this is always up to date, if you want to make sure check there yourself.
This guide assumes that you:
- Have already generated pictures with SD on your local hardware.
- Have been using Stable Diffusion for longer than a week.
- Are familiar with how LoRAs work and how they are trained.
If any of the above is not true, do not go through this guide until all those points are true for you.
Kohya-ss installation:
Ensure your GPU is of Nvidia branding, and has at least 6GB of VRAM.
Follow install instructions here: https://github.com/kohya-ss/sd-scripts
Close powershell window
Argent's setup:
Go into sd-scripts
Make three new folders in sd-scripts: .image-dir, .reg-dir, .output.
Save use_me.ps1 into the sd-scripts folder. Link to version June 11th 2023: https://files.catbox.moe/f3nri9.ps1
Edit use_me.ps1:
Change $model_dir to your ...\stable-diffusion-webui\models\Stable-diffusion folder.
Change $model to your model's filename (not your LoRA)
Set $clip_skip to model's clip skip (FluffyRock uses 1, Fluffusion uses 2. Ask others if unsure)
Preparing to train:
Go into .image-dir, make new folder of format #_CLASS, where # is the number of repeats, and CLASS is the class token used for the dataset (If unsure, just use name of concept being trained). If unsure, just set # to 1.
Put dataset (images + captions) into your #_CLASS
If you do not have a dataset prepared, I suggest looking at these threads:
ā š§tuning-projectsā pikaft e621 Posts Downloader
ā š§tuning-projectsā Dataset Curation Tool / Tag Ediā¦
Running training:
Right click on use_me.ps1 and click "run in powershell" to begin training.
If your training performs one step, then it's working fine. If it runs an error, please share a screenshot of the error here.
Once training finishes/error is shared, close powershell window.