You should check out the TouhouProject AI discord https://discord.com/invite/touhouai There are a lot of people actively training models and sharing tips in the #ai-art channel and #kohya thread under #general.
The particular way this LoRA was setup for training was an equal split of images containing her original outfit and images where she wears something else.
I had around 50 images in total.
The tags were straight from danbooru with the underscores removed.
amber (genshin impact) was at the start of every caption file and the option (--keep_tokens=1) in kohya_ss's trainer was used so in the case of an image having a very long caption that exceeds the token limit, amber's name would be kept as part of the training example when the tokens of the caption get shuffled.
The most important part of training is realizing you will likely have to retrain several times as each set of data is going to be a unique training problem. I don't think there is a single setting that would make it work well for every dataset.

After training, check the results by comparing prompts for original outfit based on your training data (so use all the words to describe it that would commonly come up in the training data, these were the words that were learned to associate with your subject).

Then test for a different generic outfit.

1. If the original outfit is lacking details and the character is incoherent, then it is undertrained.

2. If the original outfit comes out very clean but changing outfits is very hard and you start seeing original outfit details very often when you are trying for a different outfit, it is overtrained.
Fixing undertrained: you can train again at a slightly higher learning rate and/or train longer via more epochs. If you saved the training state, you can resume the last epoch and continue training. You may find that sometimes certain details of the character are undertrained compared to others. Quite often this will be shoes as many images tend to be thigh up so you lack data. With the training script in kohya_ss you can use repeats to help balance the number of full body images to other type of images by creating a folder with full body shots and increasing the repeats of the folder so when training the model will see full body images more often.

Fixing overtrained: you can train at a lower learning rate and/or cut training off early by using less repeats and epochs. You should also consider increasing the number of images of custom outfits in your training data so it learns to generalize better. Like the undertrained case, you can use repeats to increase the repeats for a folder of generic clothes.

Typical settings I start with are:
Learning Rate 5e-5 (default is 1-e4)

Aim for about 1000 steps via epoch and repeats manipulation. This ensures I finish training fairly fast so I can iterate sooner.

Figuring out how much to adjust each parameter is still a mystery that I'm trying to work out. I'm still trying to develop a method for testing and coming up with new training parameters based on simple calculations.
Good luck!

I have not. This is the commit of the web ui I am on 2725f0ac439f7e3b67858d55900e95330bbd326, from Fri Jan 13.

This is the commit I am on for the additional networks plugin 5abd8b0aedf5f282b2ebff4f43006c5cfec3fde0 from Jan 14.

If its a bug coming from the extension code you can try asking on their Github issues page.

Which model have you tried prompting for her on?
I found that she's pretty well recognized on a lot of models already without lora or text embeddings.
Something like:
hu tao \(genshin impact\), brown hair, red eyes, porkpie hat. Got a lot of her details already.
I can take a look to compare accuracy at some point.

Discussion of mpz's question can be found: https://huggingface.co/breakcore2/loras/discussions/1#63e0e4b65c6964861ebc3fd6