fingering pussy test. needs the help of other LoRAs or a specialized model to function properly.
Works well with snofs 1.1 or realism engine. (low weight 0.25 => 0.5)
Works perfectly with lustify
tested with other major LoRAs and general-purpose NSFW models, the results are not as good.
Based on my tests, I do not recommend using LoRAs or nodes that plays with the model's censorship.
It does what it's asked, except for the number of fingers inside. It's 3 most of the time.
Tags: fingering, female masturbation, pussy, 1 finger, 2 fingers, 3 fingers
weight : 0.8 => 1
Vaginal only.
Trained primarily on frontal poses.
I initially tried to create something more general (including anal and fisting), but that was a failure.
There are a few watermarks in the dataset, and I did not mask the faces. (But Works very well with an "identity" loRA ....).
Training data: 147 real photos.
The example image was generated on Krea2 Turbo FP8 using just my LoRA at a weight of 1 and Snofs at 0.4, with no other tweaks or LoRAs.
Description
FAQ
Comments (18)
i have a trained lora for this too (unreleased) and I get the same 3 finger poses as you do. even though none of my dataset has 3 finger poses. I guess the training learns that fingers go inside , but it doesnt care how many. I even carefully captioned my dataset very specifically for what finger goes where. didnt help.
Yes, same here. I carefully labeled the number of fingers in each of the images. (148 photos) , mostly "2 fingers".
@Error666 yeah. my dataset is around 200 and very varied with finger count and positions front and back. spend half a day carefully captioning each one.
I'm kinda lost on how to fix it. training unique loras is the only solution? cant be.
@QualityControl I'm definitely not an expert, and haven't tried this specifically. But I have seen some improvement on character loras by adding small portions of a dataset tagged for what you don't want, and training those images as negative weight in the overall training run. So, for example, you would have 20 images that have three fingers, tag it as "two fingers" and and those into the training set as lora weight -1. That way you get training on both ends. Teach the model what you want, and what you don't, sort of like negative conditioning. Maybe you already tried that. Just an idea.
@darkwil Interesting. I'm even less of an expert; I didn't even know that sort of thing was possible :P
I trained this version directly on Civitai. I can train locally—it takes a while, but I've already made a few custom LoRAs on Krea 2. I'll look into it with ChatGPT; it sounds interesting. Is there a UI that handles that—negative weights for specific images? I don't think AItoolkit does it, but I could be wrong.
I think adding synthetic, semi-realistic sharp data might help too, to better "lock" the hand pose.
sorry if my english is bad, i use google translate.
@Error666 I don't know about civitai, but yes aitoolkit (which is what i've been using) has it. It is on each individual dataset. Called 'LORA weight' I think. Set to 1 by default. You can use it to give different weights. Like if you want your 2-finger dataset to have weight 1, and 3-finger dataset to have weight 0.8-0.9. Basically every image is learned at that multiple. So -1.0 gives you full opposite weight. The main downside is managing a split datasets and extra steps.
@darkwil Fantastic ! Thank you! a new world is opening up to me. 😍
@darkwil That's really interesting. Do you know if there is a resource (article, video etc.) that covers this in some detail? It sounds like it could be a game-changer for datasets that are troublesome to train accurately
@bennyboy_77 I don't unfortunately. I don't even remember where I read about it and can't find it. My impression is that it works very similar to negative prompt, but at training time. That's maybe a good way to think about it. Or maybe like doing inference prompt tokens with weight. (idea 1:0.8), (idea 2: 1.8). I think of it like that. You are telling the model at training time image 1 = (fingering:1) but image 2 = (fingering:-1). In this case, you may not actually want to go full -1, because it is so similar that it will fight your training. Like if you prompt (tiny breasts:2) (huge breasts:2), those fight in opposite directions. Maybe just a small nudge negative, or even just half weight? two fingers = weight 1, three fingers = weight 0.5. Half as likely to get 3 fingers, but still moving in the direction of fingering. Just a guess. All this stuff is trial and error for me.
@darkwil Thanks for replying. I'll do a bit more digging to see if I can find any info on this (aside from the useful description you've already given). I'll post a link here if I get any further with this. Thanks again.
could you change the lora name, v1 have the same name as other Loras, creating conflict. change it for something like krea2 female masturbation
Oh right, I hadn't noticed—it's been a very long time since I last posted here. I'm going to try to change that, but for now, I haven't managed to.
In the meantime, You can just rename it.
@Error666 but it will be still working, some Loras get conflicts if are renamed by hand.
@gonfinal12 I'm not aware that any loras have problems if renamed but am happy to be corrected if you've got any examples. I routinely rename all downloaded loras to a particular format or it would be impossible to sort through them effectively. Where I have definitely seen an issue is if you download a workflow that refers to specific lora names and don't update the lora field in ComfyUI to the new name. Sometimes these can be hidden inside subgraphs which can make it hard to work out the issue.
v1 means it'll get lost amongst 100s of loras
You might get better results using a thesaurus. Some words are weighted in a way that keep you from getting to what you want, where as synonyms of those words may not be. "Putting", "Reaching" > "inserting" in this case. "Inserting" has heavy weights towards penises, then towards things.
"Fingers" have had such a problem that mentioning them is risky because it want's to weight towards seeing the entire four fingers and thumb with no part missing(inside of her). ("oh, he's mentioning fingers, make sure to show them all completely intact and in view"...)
To some degree the "fingers" issue isn't as bad with krea. Using moodymix, "her knees up. Each arm wraps around each leg to reach her vagina. Each hand is putting (4 fingers deeply inside of her vagina:1.1) on each side. Her fingers, curled inside of her vagina, tug from each side." typically just works to put 2 fingers from each hand. Which is quite a bit more than what you were going for here.
"4 fingers" is the total... ML models are stupid. Learning how to understand where they are stupid can be helpful.
I've mention the words that appears most frequently in the captioning. There is inserting or other words in the captions. I captioned it with an abliterated qwenvl , clearly indicating the theme and including the words "fingering" and "female masturbation," , with a preference for the word pussy (but there is vagina, vulva etc...) and asking to count the fingers inside her.
Then I corrected the number of fingers but kept the "literal" wording around.
But my English is very poor, so it's probably not perfect...
nice, one not so good thing thing is that they all seem face towards the camera most of the time.