A model trained on the (semi-niche??) 3D Virtual Youtuber, MonikaCinnyRoll.
Nothing can replace the real thing, so please, please go check her stuff out!
https://www.youtube.com/@MonikaCinnyRoll
https://www.twitch.tv/monikacinnyroll
Will be taken down if requested.
V2 is out!
(Thank you to the one guy who gave me feedback :D)
Increased the dataset from ~20 to 30 images.
Doubled the epochs, from 10 to 20.
Replaced the half manual/half automatic tags with fully manual tags, removed some redundant tags, and added in some more detailed tags.
It should now be able to more properly follow your prompt, like hair length, eye color, outfit, expressions, or other more INAPPROPRIATE tags.
It's also more sensitive to previously mentioned INAPPROPRIATE tags, so even the most minuscule mention of "thick" can make her a bit too big sometimes; you have to get a bit more specific with what you want in regards to curves. If you use a style from an artist that makes curvy females, then good luck trying to make her smaller!
If the eyes keep coming out weird and unevenly colored, which you can see in some of the example images, then you can remove the specific colors and keep gradient eyes, or de-emphasize the colors with square brackets (which doesn't always work). Also, if the ears and tail refuse to go away even after removing them from your prompt, you can put the tags in the negative prompt.
It should also produce 3D images much better, although I haven't fully tested for that, as that wasn't what I was looking for. MILEAGE MAY VARY!
It's not perfect, but it's about as good as I could get it with what I had (which wasn't much).
Description
Initial version
Comments (1)
Hey, this is pretty good for a first lora!
The eyes are a bit blurry though, but I'd expect it to be like that with 20 images. Adding some more helps, since character loras usually need about 30 from what I've seen. If you do not have much training data to go off of, you can crop what pictures you do have and tag for the new camera angle, like "portrait", "upper body", etc. Otherwise, decreasing the Text Encoder LR relative to the Unit LR can help if you have training data in a lot of different styles, as that can blur finer details like eyes.
I also looked at her socials, and noticed the character has really pretty sparkly purple and blue eyes in some depictions. "gradient eyes" is the proper tag for this if you want to include that for the next version.
I don't think you need to include "female" if you have 1girl in the prompts either.
Edit: Also, I think "booty shorts" should be "short shorts", and you don't need to do duplicates like "whicker markings, blue whisker markings". Just do "blue whisker markings" in the tags.