I won't lie, I struggle a lot with this style...

For use instructions: https://civitai.com/models/485413?modelVersionId=545572

Training process:

It is always the same thing, in the end only the dataset and the tagging matters.

I use the images you provided and generate some portraits with the previous LoRA, then I added the best looking ones into the dataset.

Automatic tagging with DeepDanbooru (threshold 0.5) and E621 (threshold 0.8)

Double tagging is mandatory for style, as E621 (the Pony tagging) is pretty crap for pictures without living beings.

Not using E621 is also a solution for this case.

Removing of the tags:

- canid, canine, equid => common errors during the tagging when nothing alive is recognized

- digital media and everything related to styles => they are parts of the Aenami style and must be identified as the trigger keyword

- false tags which can poison the concept, IE when I see something like "1girl, long hair" on a forest landscape without humans

Manual addings of "signature, artist name" when applicable, so it is not always generated when asked for the style, and be removed using the negative prompt.

I used the Civitai trainer for this LoRA, as the following settings aren't compatible with my graphics cards.

You can read the full settings when using this LoRA.

LoRA Rank: 32

LoRA Alpha: 32

Steps:I increased the number of steps by doubling the repeats and the number of epochs.

Styles need a longer training.

Min SMR Gamma: 0 (5)

Noise Offset: 0 (0.03)

I just don't want noise on my dataset, this would be a non sense.