So, this is my first attempt at training a LoRA for SD 3. The main idea here was to add some nice lads with noice facial hair and some stylish haircut to SD3. So it's not focused on a single toon.
It works better at weight 0.6 to 0.8 as higher than that can produce plastic/burnt effects.
I found better results with DPM ++2M SGM one and Clip Encoder f16p. But you can try setting recommend by community, as it seems to work too and may be able to give you a bigger scope than just facial portraits here. Although keeping as portraits may be good if you're looking for a more styles type of images to build your own dataset.
So, the latest epoch is working fine. Hence I'm still leaving the others: the best epochs from the batch, so you can alternate till you find the composition you're looking for.
I added full body shots, but I recommend this LoRA for Portraits/Cowboy-Shots.
Dataset was made of generated images, as we cannot use real people and I never use them anyways, as my other LoRAs are my toons which were made for my rpg story, I picked a small dataset of different lads - on purpose so that they would blend into different faces here, as this is not a single toon focused lora, then retouched them on Krita, and increased the resolution to at least 2k and 600 dpi through Mitchell.
