This was trained using a bunch of lowres colorized images from a manga. I honestly don't know what to expect when training this. To be honest this is really just an experiment because the original source material had gone through a few process before being trained so I do expect the outcome to deviate quite a bit from the source.

Her prompts:

mschni, black hair, long hair,

Her outfit:

waitress

Version 1,0 notes:

So honestly the outcome was a nicely trained model that somewhat kinda barely looked like the source. But the end result is nice actually. This should be treated as an original character as it barely looks anything like the souce material.

Even at 2.0 strength I don't see signs of oversaturation and doens't look any closer to the source material. So i guess this model training kinda failed even though the result is a really nice looking character.

Not able to test other checkpoints at the moment, but the result for prefect rocks.

Version 1.01:

Somewhat different from version 1. I'll just leave here as it is.

Version 1.02:

A much undertrained version. At strength 1 looks very similar to verison 1.0. At high strength has closest resemblance to souce material.

Additional notes:

After playing around for a bit. Base pony diffusion can generate quite interesting results at high strength. As a whole all versions still deviate quite a bit from the source material. I may retrain from scratch using a much narrower dataset to get a more consistent look.