V2:Gaaaaaaah!! *Foams from the mouth.*I had completely forgotten how awful this project was! oh the stupidity of youth. I call this the second son syndrome in which mothers seemingly forget what a pain giving birth the first time was.I burnt a bit more than a week cleaning the synthetic dataset because i was pretty much fed up with it last time. Interestingly enough fixing the dataset lowered overbaking... a lot... as in it became stupidly under baked. I spent the week progressively tweaking the learning rate and dim up the wazoo.Regardless, the only explicit benefit over the V1 is slightly less style contamination.I'll probably upload the LORA equivalent tomorrow.Seems to work ok with &lt;lora:TanetsukeV9:1&gt;, (Tanetsuke),Same poses and warnings as V1 one apply. It will likely require tweaks either in the tanetsuke tag weight or the LOCON weight depending on which Loras you are combining it with.The standing pose with no sex seems to work better with dresses might be a bias in my dataset as it is fully synthetic.V1:Well this is the fruit of my <a target="_blank" rel="ugc" href="https://civarchive.com/models/68953/fertilization-for-inpainting">inpaint LORA</a>. With that i spent several days inpainting a mostly synthetic dataset. It was kind of awful as i had almost no confidence in the result of the project. That particular LORA and then LOCON ended passing through 20+ iterations with dataset cleaning and tags pruning and modification. This one on the other hand as i spent so much time making the dataset seems to have come out ok if a bit overbaked (maybe)For recommended setting it can vary depending on the LORA mixture you are using.A suggested trigger seems pretty much useless as this is not really a stand alone LORA but if you are using without anything else you can try: &lt;lyco:Tanetsuke:.6&gt;, Tanetsuke,It seems to vary from as little as .35 to as high as 1. Al lower strengths it might no put the bubble pointing at the character belly.I'll add a SFW gallery and a NSFW one so people can see how the bubbles appear with different sex poses.In the dataset i included the following poses which i hope will be correctly displayed with their bubbles<ul><li>reverse cowgirl facing viewer</li><li>reverse cowgirl facing away from viewer</li><li>standing leg up facing viewer</li><li>doggy from behind facing away from viewer</li><li>Some prone bone position</li><li>mating press</li><li>A couple after sex poses</li></ul>Hopefully it should be flexible enough.I will add the full dataset including the raw before inpaint with the masks included. So if someone wants to train a lora and turn all the fertilized ova into penises or apples or whatever they should be able to generate a similar dataset using batch inpaint.

Impregnation for Anime NSFW (LOCON Multiple positions)

TanetsukeV9.safetensors

data.zip

v2.0

v1.0

Lora version of my <a target="_blank" rel="ugc" href="https://civarchive.com/models/77264/impregnation-for-anime-nsfw-locon-multiple-positions">Impregnation/Tanetsuke V2 LOCON</a> as some people simply prefer simple LORAs. I don't really see much difference but the LOCON seems to work a bit better with some misbehaving models likely as they export more layers.Anyway this started on an ill conceived(pun intended) idea a while ago as the other similar loras were extremely style heavy, limited in working positions and mostly overriding hairstyles as they were created with very limited datasets. So I embarked in the route of creating an all synthetic dataset as varied as possible. It was a pain in the ass and resulted in my <a target="_blank" rel="ugc" href="https://civarchive.com/models/68953/fertilization-for-inpainting">inpainting LORAs</a>. Using that I inpainted The dataset for the V1 LOCON. At that point i was a bit burn out and said good enough and tried to forget it even existed.At some point I started retraining my first LORAs and as expected i reached the point in which i once again stumbled on this... So this time I gave a manual cleaning to the synthetic dataset. The main problem with this concept is that SD and NAI have absofuckinglutely no idea of what it is and any contribution of the model is for the worse. So I had to basically sear the concepts into the lora with sheer spite. I had actually forgotten about it.Anyway, This seems to have less style contamination than the V1 LOCON, but the LORA version seems to be a bit more flaky that the V2 LOCON so you might have to push the weight of the lora above 1.I might revisit this crap for a another try in 6 months when i have forgotten how annoying it is. Seems to work ok with &lt;lora:TanetsukeV10:1&gt;, (Tanetsuke),In the dataset i included the following poses which i hope will be correctly displayed with their bubbles:<ul><li>cowgirl facing viewer</li><li>reverse cowgirl facing away from viewer</li><li>standing leg up facing viewer</li><li>doggy from behind facing away from viewer</li><li>Some prone bone position</li><li>mating press/missionary</li><li>A couple after sex poses</li><li>Standing no sex pose</li></ul>I'll add the dataset so you can check it and get an idea of what should be posible.

Files with SHA256: 24cc59c35ed7a041dfa4f339316c091f9ad8793a36dbe8b1ce78174629d98614