Thank you, though I can't really say I'm the best to learn from honestly haha.

This model was trained on 17 images, at 25 repeats for 2 epochs. so 850 total steps.

I used the Novel model for training, at 768,768 max res, using buckets.

I always edit training images to remove text, or generally improve images any way I can.

When I'm done training I run a x/y plot to determine the LoRA that works best around a strength of 0.7-1.0.

I trained this with 128x128 network size, then resized to 8. You should play around with this, not only does it drastically reduce file size, but it often softens the "burnout" effect. There's probably multiple ways to resize a LoRA, but I use the Kohya_ss tool.

I've done side-by-side comparisons where I took a 144MB, 128 network size LoRA, then resized it to 128. It reduced the size of the LoRA, but the output was nearly identical. So now I always resize the LoRA. My Ramen Girl LoRA is a good example of that.

If you use the Additional Networks extension for the Auto1111 WebUI, you can check out all the settings I used. You can load any model and then see a lot of info. I'm always looking at what others are doing.

Sometimes I train multiple LoRAs then merge them together, that also softens the "burnout" but it has mixed results and takes a bit more time.

The best advice I can give is try starting with the fewest, but highest quality images first. 10 high quality images will be drowned out by 40 lower quality images. It's better to train with 10 high quality + 5 lower quality then 10+40.

I hope I'm not just rambling and some of that's actually helpful.

There's a lot of info here https://rentry.org/59xed3 and https://rentry.org/ezlora