Hello degenerates and men of culture.
I trained this model on 40 images of high quality art of a girl choking a guy with her thighs. I was inspired by someone else's facesitting model: https://civarchive.com/models/11271/facesitting-girl-sitting-on-face
It quite reliably puts two characters in one specific pose. The model seems to be quite prone to making artifacts, so I usually lower the CFG scale or weight. It does not play nice with other LoRA's but your results may vary.
I'm releasing this for free because I'd just love to see more headscissor art, I'm really hoping someone can improve on it , but I think the results are pretty great already. Please share any images you create and make sure to tag me
Consider supporting me on my fanbox: https://arturito.fanbox.cc/
Description
FAQ
Comments (11)
can u make choke hold lora next
Really wish there was something like this for bodyscissors -- criminally underrepresented hold, considering how impressive it is for a woman to crush an entire guy rather than just his neck.
Just wondering what the process was to train these, trying to do something similar, I do have the ability to utilize Daz to render a specific pose at multiple angles (with maybe slight alterations and camera angle changes), but I haven't gotten exactly what keywords to train the lora yet to get even close to the results I get with this download. Just wondering if you had even a bit of the workflow in how you tagged and created it so I can do something similar.
Sure, love to help out.
I've updated this model with a better one, so my process changed from this one. Basically, you want to have your "trigger word", which you keep as the first tag. Making sure you do "keep tokens 1" so when you shuffle tags it always remembers that word. Then for the tagging, you basically want to describe everything in the image except for the essence of your desired concept. So for example clothes, expression, characters, image quality, etc... but avoiding describing the pose or anything essential to the pose.
If you're doing it on anime style, it's much easier, since you can use autotaggers that work much better than BLIP. Just make sure to take out anything that describes what you want to train on.
And for complicated poses like this, especially with two people, you want to train on MANY images at a low learning rate for a long time. I only started to get really good results at 300 images, but you many want more.
Daz is a great idea, just make sure the images you use are as varied as possible in terms of lighting, clothing, characters etc.... with your pose being the same throughout the images.
@Arturito Sounds good, did you do anything tag wise related to who was in it? Like for the images, 2girls, or 1girl 1boy, also position wise? I planned on using dreamscape as the model to train from, to use some Booru Dataset manager for consistent tagging. Starting with a classic straight headscissor from the front (not sure how many different angles positions you had, if I should include them all, or just focus on a specific camera position).
Thanks for the help, also noticed on your reverse headscissor one you mention there are two versions, but I don't seem to be able to find the original. I have the one from here, https://civitai.com/models/111958/reverse-headscissor?modelVersionId=120847
I have the lora file <lora:animrevscs> but I don't seem to be able to find the lora for <lora:revscs_FKV2>
@realityman0628477 If you're working on a headscissor model, I'd love to collaborate. Maybe I can take a look at your dataset.
Not entirely sure if having more camera angles helps or not. I'd like to think it does, but only if you have a good amount of images.
Theres a different tab for the other model:https://civitai.com/models/111958?modelVersionId=120844
@Arturito Yea, I'll send you what I have, I ran it off 350 images this first time here (these are live models not drawings or renders at moment), and I'm experimenting on a few different things. Couple of things that I'm starting with, the checkpoint, whether to use the standard 1.5 or use something like Dreambooth (to utilize the tags and flexibility). I set learning rate to 0.0005 and just experimenting with number of repeats set on the folder and the epochs. Ideally it seems like when you do a specific pose like this, I feel like it should almost be pretty rigid as opposed to what you would want from another training, but maybe I'm wrong on that.
Tagging wise I ran it through an auto tagger and removed everything related to the pose (the positioning, actions, etc) and left everything descriptive in like what clothes are being worn or the background. I'm still not sure whether to leave the tag in for if there are 2girls in the photo for example.
I noticed in yours you added tags in most of your images, like "head between thighs" and "headscissors"... which I don't think are in the dataset tags, but instead to help guide the AI to understand what it should be drawing via the tags and also via the trained word (i used xypclashs)
Plan on doing a reverse headscissors attempt, a victory pose set, and a reverse facesit as well, I'm going through my images and getting examples now for those.
@realityman0628477 Do you have discord?
@Arturito I haven't been on in a long time, I'll see if I can remember the credentials, in the mean time first few runs were okayish, Id like to see what you typically tagged on your images (I know some of the comments on the loras you mentioned they were not tagged correctly for about 80% of the images) to compare what I did (here is an example of one of my tags, trying to tag everything about the photo other than the scissor hold:
xpyclashs, breasts, smile, open mouth, blue eyes, multiple girls, brown hair, jewelry, sitting, underwear, panties, closed eyes, lying, shorts, barefoot, on back, bra, feet, lips, legs, short shorts, toes, soles, denim, black bra, lace, denim shorts, realistic, anklet, dirty, dirty feet)
. I also tried to cut down on buckets by resizing into 3 or 4 of them, trying to give the AI a better idea of the images. Also working on playing with the number of steps/epochs... seeming to go way past what I did for a few character loras I tried that came out pretty good for my expectations.
@realityman0628477 I'd love to take a look. If there's a way we can DM I can send you my dataset for you to look at too.
Could you upload to tensor art please?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.