Type | |
Stats | 1,157 |
Reviews | (201) |
Published | Jul 11, 2023 |
Base Model | |
Trigger Words | marle barrock |
Training Images | Download |
Hash | AutoV2 6C90828080 |
Marle Barrock from Reincarnation Colosseum for the approx three people who know who it is. And for the rest - no, you can't fix her.
(LyCORIS version here)
Prompt: 1girl, marle barrock.
Optional prompts: "ringed yellow eyes", "muscular female" and "dark-skinned female"
Standard outfit: bandaged arms, bandaged legs, elbow gloves, thighhighs, tabard, pelvic curtain.
For more prompt tips see the sample images or the zip of training tags.
The dataset for this was rough and I had to do a lot of shenanigans and retraining to get a working LoRA.
Still, it's not perfect, due to the lack of useable images and the weird outfit lacking good tagging it can bleed over but just use the outfit tags in negatives if that happens.
You can also drop "marle barrock" entirely in some cases and still get decent results if the style is doesn't play nice with your model.
The below information is for the first version. 1.1 should be more flexible and easier to work with.
Trained on NAI it gives results closest to the manga, especially when used with "greyscale" and "monochrome". "hatching (texture)" was also autotagged but it's a fairly weak one so don't expect much of it, you probably need to escape the parenthesis for it to work - "hatching \(texture\)"
AOM2 gives as usual a more semi-realistic look, "3D" and "realistic" along with the above tags should bring it a bit closer to the source material though.
Don't forget to review and upload your own images, it's always fun to see what you're creating!
MISC
Trained on NAI. Dataset of 41 images + 3 selfgenned.
All sample images are raw txt2img, slightly cherrypicked, no inpainting done.Two or three promo art shots shows Marle has dark red/auburn hair and yellow eyes but those wasn't enough to train on so the dataset is entirely in the manga greyscale/monochrome form.
The training was done on a small subset of the artists images, as such the output of this LoRA is not an accurate representation neither in quality or style of the source material. Go support the artist ZUNTA if you want the good stuff.