ghost
Sign In

Oatd Style

18
138
161
4
Updated: Jun 19, 2024
styleanimepinupswomancartoon
Verified:
SafeTensor
Type
LoRA
Stats
138
161
Reviews
Published
Jun 19, 2024
Base Model
Pony
Training
Steps: 10,260
Epochs: 20
Usage Tips
Clip Skip: 1
Strength: 0.8
Training Images
Download
Hash
AutoV2
3228D0F3F5
default creator card background decoration
Bronze Writer Badge
urusan's Avatar
urusan

This is a fairly general purpose pinup (single woman subject) LoRA. It produces a hybrid between western cartoon and anime styles. It was derived from another (currently unreleased) Pony model trained on a single image (warning: NSFW) from the artist Woot. I then used AutismMix to generate examples without the subject of that image and liked the resulting style. It's not really Woot's style, even though it includes certain elements (most notably the eyes), so I came up with the random new name "Oatd".

In addition, I used a (new?) technique in training: caption augmentation. I prepared 3 "levels" of captioning detail and trained all three together. The results were better than I imagined. The model is very good at any level of captioning detail. The included training data is in 3 folders corresponding to the captioning detail levels and should be folded together before training.

While the model should be able to handle a wide variety of prompting styles (though Booru style prompting is recommended), and there's no specific trigger word, certain prompts are especially powerful due to their use in the low captioning style and thus they should generally be included for best results.

woman - All of the training data has this caption, so it's recommended to include it for best results

from front/from side/from behind - Pick one of these options to ensure you get the right angle, it broadly prefers "from front" by default

indoors/outdoors/simple background - Pick one of the options to get the right overall setting, it prefers "simple background" by default, you can alternatively just be more specific about your setting and it should start to pick up what you mean.

The low captioning style also includes basic information about attire, body positioning, and body shape, so those captions are recommended as well. You can leave out any of these captions and it should perform fine.

On the other end of the spectrum, it's also well trained on keeping colors separated in the high captioning style so it does well at keeping colors of clothing/hair/eyes/etc. specific to that thing.

One last notable caption is face focus which will give a close up of the character's face. Useful for inpainting!

The model itself is very general purpose and should be able to make SFW or NSFW depending on how you prompt it. Some of the included training data is somewhat NSFW, including nudity and absurd female proportions, but it doesn't contain sex. By default it ranges from SFW to mildly risque.

1024x1024 or 1024x1053 is reccomended, but hi-res fix to 1.5 usually helps.