[ANIMA / Illustrious / NAIv] NEW re-trained under ANIMA/IL
Use
redsage, short hair, white hair, red hair, multicolored hair, hair over one eye, red eyes, black dress
or
bluesage, short hair, white hair, blue hair, multicolored hair, hair over one eye, blue eyes, white dress
to summon her
Supplement tag to define her appearance: digital dissolve(for glitch effect), pantyhose(if don't want her flesh legs appear), mary jane
* weight white hair more than red/blue hair tag if her hair gets more color than the white.
[NEW v6!] : Train with more source image, balance between child and adult apperances.
V5 : More source images, very fine tuned caption for more flexible prompting, remove her glitching effect in source images that mess up algorithm
v4 : More training source images and better caption management, you can summon her with just only sageai
v3 : Train with new method, react better whtn prompts.
v2 : Uses only SFW pic to train, use better captioner
(v6) NEW!
Use sthsage to summon her, use short hair, white hair, multicolored hair, hair over one eye, red eyes to form her head, use black dress, black pantyhose to summon her default outfit
(V5)
Use sageai to summon her, use black dress, red gloves, red legwear, to summon her attire and red glitching covering her body.
(v4)
SAGE (alt. adult ver.) from Sonic Frontiers
Uses sageai to summon her, just that.
Old version. (v1-v3)
Uses sageai to summon her, use red eyes, short hair, hair over one eye, white hair, red hair, multicolored hair to set her head, uses black dress to give her signature dress.
Disclaimer: The characters were depicted as 18+ years old and the training data used to make LoRA is an adultified version of SAGE.
Description
very fine-tuned caption and better image source, 50 in total.
FAQ
Comments (3)
Learning how to create LoRAs myself right now, and I have to ask, how on Earth are you able to get such fantastic backgrounds? The few LoRAs I've made have the character just fine, but the backgrounds are awful. And how do you get it to get Sage just with a single tag?
When you train it, do you just train the images on "SageAI" alone? Do you train backgrounds separately or shop Sage into good ones and trick the AI? I'm intrigued.
The secret sauce of training LoRA is in fact...captioning.
Caption is the .txt file with the same name as the images that contain the description or tag that the picture represents. For example, If I use WD1.4 automatic image tagger in Kohya on this Sage official render (https://sonic.fandom.com/wiki/Sage?file=Sage.png), I will get this.
"
1girl, solo, looking at viewer, short hair, simple background, red eyes, long sleeves, dress, closed mouth, standing, full body, white hair, wide sleeves, hair over one eye, black dress, glowing, transparent background
"
This is what I mean about the caption, it describes an image for LoRA trainers to understand what is what. If I want it to learn "sageai" I just add "sageai" in caption like this...
"
sageai, 1girl, solo, looking at viewer, short hair, simple background, red eyes, long sleeves, dress, closed mouth, standing, full body, white hair, wide sleeves, hair over one eye, black dress, glowing, transparent background
"
This will give the most flexible LoRA but use more tags to generate Sage since you have to add these in prompts when generating her.
However, if you want one tag to describe Sage and generate her with that only tag, you must modify the caption by removing any tag that is associated with Sage, that way all the removed tags associated with Sage will merge into "sageai" instead.
"
sageai, looking at viewer, simple background, closed mouth, standing, full body, glowing, transparent background
"
As you can see, we leave background, expression, camera position, etc intact since we want this aspect to be a variable, not "sageai". That answers the question of why your LoRA has a bad background because AI thought the background was part of "sageai" while training
This captioning style will make LoRA too easy to use, with just a few words. However, this will make a LoRA very stiff, You can't change hair color, clothing, etc since those were baked in "sageai" already.
@Smez That explains quite a bit. Thank you!