Hello!

Welcome to one of my weirdest experiments I have ever done. Dual LORA [Character+Splitted Style].

But, this wasn't Sagat from Street fighter? Yes and no.

Short story, this LORA was not trained under normal circumstances and it can be used to create Sagat in a very specific style on a lot of models and use that line style to apply on any other concepts but using Lykon's AnyLora (primary).

So... why not a normal LORA?
********************************
Cause if you put just one Sagat image with his eyepatch (captioned or not), it blows the face and the body in the training and the inferencing. Tried 15 trainings this week, nothing.

What I did? Hands at work in Photoshop (time without drawing lol) and created several illustration and trained from there, I have added a pureref chart in the end of images to explain this best. Basically I had to draw, again (and using AI on my advance to speed up this), lol. This way I get cleaner images to train and without his eyepatch. The scar is other story and it reflects it more or less depending on the LORA's strength.

That's why in the training it got my lining style (splitted training) plus the character. So we can say this is an stylization or a fan art (and I'm not gonna do this anymore except when the AI's internal understanding limits the trainings).

Token is Sagat and for the style just not use Sagat and drop to 0.7 (recommended) the strength, Prompt anything you want. Easy.

Recommendations
*********************

- Models to inference: If you just want Sagat you can use several (AnyLora, RevAnimated, Colorful..) but if you want just the style is extremely recommended and I can't say this will work in other models (cause maybe conflicts), Anylora.

- Strenghts: 0.8-1 for Sagat prompting and 0.6-0.7 for style in AnyLora.

This LORA is good for:
************************
- Sagat half body, fullbody, action poses and very low profile portraits.
- It can do backgrounds till a certain level, if you feel like you're not getting them drop Sagat from 1 to 0.7/0.8.
- Anime/manga/lineart and some semi realistic/concept styles depending on models.
- Lineart style punching the contrast (under the circumstances of AnyLora explained above).

Sorry for messing so much this time and hope you don't find this extremely complex.
Thanks to Lykon for the Anylora's, didn't know it was so powerful when training (specially different images).

Feel free to support us at:
https://www.patreon.com/user?u=90661954

Thanks to all!