Hi,

I used 108 images. In fact, I worked on only 54 images. I just duplicated all of them and then cropped them differently. I have heard that this can reduce overfitting, but I ain't exactly sure how much it helps tho.

I didn't use any fan arts. There are too few of them, so I decided to train without them.

I trained on NAI.

I see.

Actually, I had already upscaled my images before training them with R-ESRGAN 4x+ Anime6B, but usually details are degraded a bit. AnimeSharp is better option if I want to preserve more details, but I haven't tried using it on my dataset yet. So far, I have only used it on my generations.

@testaiproton 

Hello, thank you for your comments.

I'm not sure if using the official character tag is a good idea as I think it may negatively affect the model. For instance, if the official tag includes artworks with multiple characters, it could potentially confuse the AI. However, this is just my speculation.

I haven't checked for any recent LoRA guides, but you can try reading the followings:
https://civitai.com/models/22530/guide-make-your-own-loras-easy-and-free
https://rentry.co/59xed3
https://rentry.org/LazyTrainingGuide
https://rentry.org/LazyDAdaptationGuide
I can't confirm whether the mentioned training parameters are good or not. You'll need to experiment yourself. Additionally, you can examine the metadata of models created by popular model trainers to find out what parameters they use.

I'm not sure if my model looks good or not. Based on the sample images, it seems that many others, like Mr LittleJelly above, have done a better job than I have.

Btw, all my models are just my own experiments. They were rarely trained with the same parameters, and the datasets were often modified in various ways. For example, in the past, I tried painting the background color of some images to reduce background overfitting. I also experimented with mosaic censoring and blurring to hide unwanted characters or details when cropping them out wasn't possible. However, recently I've been too lazy to do these. Based on my experiments with two rare (hentai) characters, I realized that I probably didn't need too many images in my dataset. Quality over quantity. Nevertheless, you'll need to do your own tests. Another experiment I did was to remove the anime art style from some of my models using certain techniques, although the results may not look very good. There were other experiments I did, but I don't remember them all, and it is boring.