Type | |
Stats | 80 21 4 |
Reviews | (14) |
Published | Jan 25, 2025 |
Base Model | |
Trigger Words | makoto_minamike makochan_minamike makoto (minami-ke), minami-ke blazer, red tie, randoseru pink dress |
Hash | AutoV2 50F1A63B3F |
, watashi ni tenshi ga maiorita!, portrait, _3, b s-3698372704.jpeg)
A LORA for Makoto / Mako-chan (i.e. trap version) from Minami-ke.
This is a little experiment of mine, plus a project I want to do.
The project is to try to create LORAs for as many characters from older anime as I can. I think AI can help "preserve" them so to speak, to be able to generate them again. When I was gathering images, I noticed many blogs I used to read back in the 2000s, while they exist, wouldn't even load the screenshots anymore, probably the image servers long migrated or gone - which makes me want to train LORAs so I can always recover these characters and animes. Who know how AI will evolve in the future.
The experiment was related to training a LORA with two related but different concepts and trigger words. In this case, makoto_minamike is the main trigger, and makochan_minamike the secondary one when crossdressing.
My first attempt used makoto and mako-chan and didn't work too well.
This second LORA has some quirks, but I decided to post it. The quirks are probably related to the training data:
It looks very anime-like and very upscaled. The reason is simple. The BD version looks upscaled over the 2008-ish original, and I also used Topaz to upscale some other images. However, prompting some different artists or longer prompts seems to dissipate this effect
It surprisingly tends to generate him either from behind or in close-ups. The reason for the close-ups is that anime screenshots being horizontal tend to focus on the face. The reason for images from behind is that maybe I used too many shots from behind or from odd angles to try to give the training set more variety but maybe it was too much. Using facing viewer, etc helps but it's not ideal
However, the second version I trained seems much better at differentiating Mako-chan and non Mako-chan. In particular, without having to include the tag, Makoto (male clothes version) gets quite consistently his ahoge, while Mako-chan doesn't, since his hair is down. In my first version, he always got the hair ornament, so I tagged it for the second training I did, but this means it has to be tagged for the Mako-chan version or there are chances you won't get it. Ideally, Makochan_minamike should have captured this nuance too.
Another issue is eye color. I had no idea what to do. In the first season he had brown color, then it seemed to get black, and by later seasons he had green color. I didn't tag eye color, and in the end the AI has no idea what to do with that, so it's better to tag it as desired.
The model seems to work no WAI NSFW and Cat Tower, but NTR mix seems broken.
All in all, as usual posting it as it is. It has its quirks, and will probably re-evaluate the training data, tags and try again some day, but it's an interesting experiment, and I wish to find more characters from older anime to bring back to "life".