V2.0: Retrained so it's roughly the same number of steps but now 50% on a large dataset and 50% on a smaller cherry-picked one. Not strictly better than V1.0, but should affect the original prompts less than before.
Pretty simple Lora that adds tape gags to characters when prompted. I made this and a couple earlier versions a while ago before there were others on here, but figured it was worth uploading anyways. I've mainly been using it on some bondage mixes using AOM and other anime-based models with reasonable results, though it still microwaves the image in certain models and with too many other Loras due to over-fitting. Also seems to be way worse on average outside of bust shots, so the preview pics are pretty boring as a result. I'd like to improve that, so feedback appreciated!
Can also be used through in-painting, but I find it rarely matches the face well in those cases. Tags such as "gag", "gagged", etc. were removed from the training set so only the trigger words would be necessary.
Recommended weight: 1.0 on bondage models, ~0.5-0.75 otherwise.
Recommended CFG: ~6-7
Models used are all some mix based off of bondage-v11: https://civarchive.com/models/34777/bondage-v11
Definitely check it out!
Specifically, the Meina model is:
meina-pro-mistoon-hll3.safetensors + bondage-v11 + v1-5-pruned-emaonly
Multiplier: 0.5
Add difference, copy config from A, B, or C
Description
Trained on animefull-final-pruned.
Experimented with mixing datasets for this one, so trained for around 3000 steps on a large dataset of ~1100 images then another 3000 on a cherry-picked subset of ~90 images. Fully training on the large set seems to get the tape placement right often enough but tends to mess up the tape itself while only using the small set (v1.0) gets the look better but noticeably affects how it interacts with the rest of the image. This version seems to be a middle ground the appears more stable but occasionally worse than v1.0