Ovary/Uterine/Vaginal Prolapse (Moss Version)

training with danbooru tag prompt, the tags are:

ovary prolapse, uterine prolapse, vaginal prolapse, ovary, fallopian tubes, disembodied ovary

if the ovary is detached with uterus, add disembodied ovary in negtive prompt.

== V2.preview4 ==

• we skip preview 2/3 release and go to the preview 4. Now we know more about the LoRA training in Moss Image, if future test training has no more improvement, this will be the last version of ovary prolapse LoRA for Moss Image.

• The major reason we stop training is the lacking of the tagged training data. If you have more clear free uncensored tagged training data, we can have another trying. But all those data must be able to released in public repository.

Add a few more concepts in the training data, espically the layout one.

existing concepts in Moss Image:

cervix, ovary, fallopian tubes

these concept is overrided in the LoRA, if you want one with prolapsed one, you have to inpaint in the generated image.

new concepts:

raised prolapsed vagina, raised prolapsed uterus

noticed that the raised version using n. form not the adj. form words. this is designed to reduce mixture of concepts. if you can't generate the shape as you wish, you can try using these tags instead of using NL.

== V2.preview1 ==

Add layout prompt in training data.

the controlling is better in tag + nl.

in layout prompt, the bbox works if the position is nature. or it will obey one of them and ignore others. you can try this:

{"summary":{"caption":"masterpiece, best quality, highres,general, sensitive, nsfw, explicit, visual novel cg, Japanese galgame screenshot, barbara_\(genshin_impact\), sitting, looking at viewer, pussy, vaginal_prolapse, ovary_prolapse, ovary, fallopian_tubes, pussy_juices, fullbody, wide_shot, simple_background, white_background,"},"items":[{"label":"object","caption":"vaginal prolapse","bbox":[56,-29,138,42]},{"label":"object","caption":"uterine prolapse","bbox":[104,-30,171,44]},{"label":"object","caption":"ovary prolapse","bbox":[153,-38,226,-9]},{"label":"object","caption":"ovary prolapse","bbox":[170,5,265,35]}],"bbox_coord_space":"abs_centralized","bbox_format":"yxyx"}

this lora is still under experiment. maybe some problems in prompt understanding (and some of them may never be ableto solved)

== V1.0 ==

Using the dataset what is used to train my Anima version ovary prolapse LoRA/Lycoris.

No Natural Language/Layout Json Prompt add now, but it works under these prompt style. Layout Json Prompt is not very stable for now.

Add some relative position description to make the target more nature. e.g.

ovary prolapse on ground

There is some style side effect, you may need more style prompt to suppress the side effect.

The research on moss lora is still under progress, I can't expain much about how to write the prompt effectively or how to overcome some problem in generation. you should do your own research on prompt writing.