Noobai doesn't need a grea lora, has good knowledge of her already. When Vpred 1.0 releases i'll think about it and see how my dataset as a lora affects noobai's knowledge
Illustrious | Tofu version of my first LoRA because I like Grea and didn’t see one for SDXL/Pony yet, at the time. The Tofu version works fine with AnimagineXL, i think.
Tags you should consider adding in positive/negative:
See sidebar first for anime default outfit and anime bikini outfit tags;
P: It's a requirement to add dragon tail in the prompt (see "About this version" for info)
P: HAIRSTYLE: short hair or long hair, hair over breasts also hair censorhelps.
And: ahoge if you want it; looks cute :>
P: anime screencap to achieve anime style
N: detached tail, multiple tails, missing wings seems to help in general
N: shoulder straps when prompting high waist skirt (depends on model but they appear often from my testing because of the high-waist skirt)
N: breast pocket to prevent said concept form appearing on the shirt/nightgown
N: lace-trimmed legwear or (P)plain thighhighsif you encounter frilled/lace-trimmed thigh highs although from my testing they weren't nearly as bad as in v2 (i actually forgot to tag them in v3 so they got trained in, my bad)
P/N: high-waist remove it or add it based on your preferences. Controls skirt somewhat
P/N: lace-trimmed bikini (weak) for the cyan-ish "fins"/lace-trim from the anime bikini
Note: o-ring bikini might be unstable in certain scenes. If a new version is gonna be trained it'll be fixed in there, probably.
Flexibility (Colors/Concepts/Styles):
Overall the LoRA should be pretty flexible. As in changing hair or eye color or style shouldn't be a problem although sometimes you need to add more weight or put the default color in negative when a color change isnt working. The high-waist skirt from default outfit seems to need extra work with color change. The wings seem to be trickier ("blue wings" for example made them completely blue but sometimes spilled over to the tail and "black wings" just kept the membranes at default color).
Color change example (Black themed Grea) copied from pony:
Here is an example of positive and negative tags i used for the black clothed/scaled grea (for pony version) (this was just a quick test to see if it's possible, i don't know if this is too many tags or not but it worked including the side ponytail hairstyle):
P: black horns, black wings, black dragon tail, black scales, black hair, dark purple eyes, black shirt, black high-waist, black pleated skirt, black plaid skirt
N: red hair, purple hair, red plaid skirt, red high-waist, red pleated skirt, blue skirt
For more info and technical info, check "About this version".
To support me without tipping Buzz, use the "Add post" button above the gallery (instead of "Post images" from the "Create" dropdown) if you want to publish images made with my LoRA. Each post earns me 50 Buzz. You can do this for other creators too. or u can tip me uwu 🥺
Description
This is mostly a test, use the pony or IL version of grea for actual use
This version is trained on the Tofu anime base model by Minthybasis and is supposed to work with it. I tested it with AnimagineXL v3.1 and it worked fine too btw. Might work with other SDXL anime models too but I don't have any to test. I'll have a version trained specifically with animagine to see if that gives better results - that is, if i don't forget it or wasn't too lazy to do it.
Issues:
Tail is rather unstable, hoping for new version of Tofu
Back shots are harder because of that, tail base on back often looks out of place, some high detail styles may fix that
There are no differences in training or dataset over v3.0 PONY besides one setting i think (which i think was Min SNR Gamma set from recommended 5 to 1 per this guide, which might be the reason why this version is ~18mb larger than the pony version although not sure, interesting though). I'm hallucinating or something, only the upcoming illustrious version has slightly different settings.
The following is the "About this version" changleog text from the PONY v3.0:
Cleaned the dataset and tags a LOT for a better result (mainly stability across concepts/outfits) and it is now easier/more efficient to prompt an outfit through less tags needed to do so.
Important: see the model description for tags to include or exclude, other than those, do that as usual with objects you (don't) want in your image
Multiple outfits available! (See sidebar info trigger words)
Main: Default (default white shirt) and Bikini (beach episode, o-ring might have issues, if so just remove any occurance of it in your prompt)
Extra: Nightgown, Winter coat with capelet
Both short hair and long hair Grea should be pretty stable from possibly all views, have fun!
Long hair style is much more stable now as well as hair over breasts and hair censor.
Why dragon tail is a requirement but other tags like dragon wings/horns isn't:
dragon tail in prompt is a MUST if you want it in the image because I tagged it in the dataset since without that tag the tail would often merge with clothes or just not appear at all even with it prompted, which would lead to several more issues such as the likeness being gone.
This is only a requirement for dragon tail and does not apply to the horns, wings or hair/eye color. To change those just put whatever you don't want in negative and to change eye color simply prompting "blue eyes" or similar.
Hair style/color change may require higher weight or red/purple hair in negative.
Technical stuff:
The model was trained with kohya_ss Network Dim 32; alpha 16. The LoRA was resized to rank 8.
If you are interested in the settings I used (if you know about training loras, please tell me if there's anything to be improved!): https://files.catbox.moe/m1klvs.json
Quick note to that: I used prodigy and 6100 steps. the reason it didn't overtrain is probably because of scale weight norms is set to 5.
note if it matters: I use kohya_ss with pytorch 2.4.1+cu124 and xformers==0.0.28.post1