#1: Lack of screen time isn't an issue if the chara can be seen nicely from the different angles. Actually I made about 30 screencaps of her just from episode 4. Another trick is to have just her on the dataset images, so I crop out other charas from it.

#2: Another important aspect is that I take my time adjusting tags, not always tho. The tagger doesn't always add everything.

#3: I usually do 3 or 5 repeats per outfit but with Rory Mercury I learned that it's not necessarily good to have outfits in separate folders because with first Lora I couldn't get rid of her "animal ears ribbon" no matter what I put into negatives. So I smashed everything into one including the ones where she doesn't have ribbon on the head. And again I tagged everything properly.

#4: 3 repeats with 30+ images, 5 repeats with less than that.

#5: Kohya settings...

I can't really tell you my secret, because there is no secret. I experiment all the time. The only thing I can tell you is that I train my LoRAs on 768,768 resolution and not 512,512 to keep as much details as possible with learning rate 0.0001. The rest of the settings is pretty much basic as in almost every guide. Oh yeah and I don't do dim 128 alpha 64 anymore. D32 A18 seems like plenty enough.

Also I train the AI on leaked NAI pruned model.

That helps greatly compared to training it on models like Anything-3 or others. The results of those are not even comparable.

For generated images I use ClearVae or automatic.

I usually generate 6 images in a batch. Regen if I don't like any of them. Also depending on a scene and details I want, I again switch between 512,512 and 768,768 resolution with high-res fix and 4x-AnimeSharp upscaler, denoising 0.6. Send the generated image to img2img, denoising to 0.1. Let it run, it will do color correction, then send it to Extra and upscale it to the resolution you want with the same upscaler as in txt2img.

Done.

It's not a rocket science but it takes a lot of time and experimenting. As I said, depending on a scene and results I want these parameters are not exactly set in stone.

@zetsubou Dataset Folder:
3_Yuzuriha_Kunoichi
5_Yuzuriha_Bikini
5_Yuzuriha_Towel
Just an example.
However I stopped doing this and I put everything into one folder and focus more on tags.
If one outfit has only one hairstyle, and there's a different one in another folder, it's hard to get that 2nd hairstyle for the 1st outfit.
So my advice is, experiment by yourself what works for you.
I usually need to teach AI 2-3 times, to figure out the golden ratio for folders and other settings.

That's rather good choice I have to say. 👌 I'll put her on my list.

@Xatro165 generic is good, I prefer anime artstyle as a source