>>43405405I'll try to do a quick retard writeup based on my very quick and scuffed lora making so far, I really have no clue about training rates, epoch and repeat counts, regularization images and so on. Just a bare minimum "what you need to do to teach the AI something with like two dozen images"
I used this lora training guide here:
https://rentry.org/lora_trainI installed this:
https://github.com/kohya-ss/sd-scripts - git pull it into a new folder like you did with the webui
Grab the training script here:
https://github.com/derrian-distro/LoRA_Easy_Training_Scripts and toss it into the folder where you installed sd-scripts - I recommend the popup script, super easy to use.
That, plus the NAI.ckpt, plus some training data is all you need to get started. The "how to use this" part of the linked lora training guide is pretty much all you need to know. Be sure to follow the pictured folder structure for your training data, even if you just have one folder to train on - the script is pointed at the top folder, and you have a properly named subfolder with your training data, even if you just have one folder for images. Otherwise you just follow along with default settings on the popups, I usually do 10-12 epochs and 10 repeats on my one folder, but you can try and experiment with these.
For training data, I've grabbed at least a dozen images and ran them through the WD autotagger extension in webui, maybe make it add an extra 'activation tag' if you think your lora needs one, remove an unwanted tag and so on.
Once you get the hang of the janky popup script, it's really super straightforward, I'd like to learn more about optimal image selection, captioning, training rates, repeats, epochs and so on but just installing the sd-scripts, the popup script and learning to use it is all you need to get started with lora making.
