A LoRA that depicts the complex body motion and fabric interaction of a woman removing her panties while lying down
The base H3 model and even NSFW merge checkpoints really struggle with this motion so I thought I'd attempt to correct this, so you can have your character reveal that last stage in the "female is receptive" signalling beautifully and smoothly.
MINIMAX H3 VERSION (NEW!)
Trained on the H3 base model (Int8 convrot) on my RTX5090 using ai-toolkit, Steps 0-6000 on 23 variable length (5-15s) clips of real life girls lying on their backs pulling off their panties. To fit the much longer than usual clips into my 32GB VRAM I halved the fps down to 12, and ran a 320x320 resolution.
Usage Notes:
For most predictable results use with the H3 Minimax base FL2VA model, although it will play nice with NSFW checkpoints such as 10Eros.
For Ref2VA use a hybrid checkpoint such as 10Eros's as this LoRA doesn't perform well with the base Ref2VA model.
10-12 seconds is about optimum length, as less than that and the girl seems "hurried", more than that and some additional unwanted movements can enter in.
Strength 1.0 has been used in all testing and showcase videos without artefacts.
As always with H3, tight and detailed prompt structure is key. See my showcase videos for examples.
Thank you to the community for your YB during Early Access. Goal hit in only 12 hours! If you like my work and wish to support me, then please consider signing up for a paid sub to my riotmodels page where you get to see the full scenes of the Seven Sisters doing their thing
Description
Initial H3 Version. Trained only on women lying on their backs pulling off panties.
FAQ
Comments (27)
Base model says Krea.
Just pointing that out in case it causes issues.
And as someone who has been trying to properly take off panties in minimax, its... been a challenge, whooweee.
Good catch. It's now corrected.
And yeah, I kept running into issues with the very important task of my characters pulling off this final intimate layer of clothing so decided to remedy that!
@Generation3dX This, and from any pose > facesitting have been my bane these last few weeks. Currently "rendering" out a whole load of other videos before I'll re-attempt the undressing, though.
@FairyDollFetishist facesitting should be possible with the base H3 model, right? I haven't tried it yet. But due to the limited visibility angles, you can't really show genital/tongue contact easily, so gross hip grinding motion, say, from the side is perhaps the best you can hope for. Also the other character's face is hidden and so I think one loses that golden character-to-character interaction which is so desired.
@Generation3dX The action itself works well, especially with a reference image. Eros eats out like it's breakfast. Getting a character into that position from any other one, though... After two days of prompting, I couldn't get it right even once.
@FairyDollFetishist if you're referring to the difficult of the seed image to get that position, yeah, I get it, especially with Krea2. Pony has a better chance. Have you used the Krea2 control net?
how much time did it took to train? just out of curiosity
About 16 hours in all.
Nice, next can we get a she sells queefs and fots in a jar on onlyfans lora?
i see you.
good job，but I'm looking forward to v2
Thanks. What improvements/additions would you like to see?
@Generation3dX taking off panties, jeans, skirts while standing, taking off a top, a shirt, etc. Minimax has huge problems with all of this.
@redsnail I never had any issue for taking off to knees shorts, pants, panties, etc...with H3. compared to Wan22, or LTX2.3 2.5, H3 does it almost perfectly.
but as always with H3, you need a good prompt.
If you still have hard time doing it, use the Penis reveal lora, and all your issues will be fixed for standing girl.
@Pat3dx Yeah standing is covered already mostly fine, which is why I didn't invest time in training that further. The panties start as worn then are pulled down below the knees fine and fall to the floor or out of shot. Very different matter, though if she's on her back. The H3 model (base or even NSFW checkpoints like 10Eros) cause her to skip her knees sometimes and definitely the feet, so it breaks immersion and renders the whole generation useless. I believe my LoRA fixes that key movement phase below the knees.
@redsnail the problem is there are so many different types of clothing, and each one has it's own pattern of removal, like bras unfastened from the back, t-shirts or tops pulled up over the head, blouses unbuttoned from the front, dresses unzipped from the back then falling down. Each one is a completely different motion and would require a separate layer of training. Because of this diversity, this is why all models including the excellent H3 model, struggle with it, as it's just tons of training. Still it's something to consider as a future project.
@Generation3dX Also gotta keep in mind that Civitai is pretty strict with loras that can "nudify" fully-dressed women, so you would definitely be playing with fire depending on the clothing you're training for.
@TwoMoreLurker hmm I hadn't even thought of that. But yes, I can imagine such a LoRA being used for unethical purposes in using a base photo of a celeb or such. Might be worth avoiding for that reason alone.
A lora that actually makes sense! Especially in doggy H3 has problems with that.
Great. When you mean doggy, do you mean on all fours / tabletop position?
@Generation3dX On all fours yeah, but you trained the lora only in missionary position? Maybe you can include that if a v2 comes?
I thought about it. Some of the RL source vids I was scanning for training clips had the girl on all fours pulling them off one knee/foot at a time, but they are rare. The position is a bit awkward and unusual. But it's perfectly trainable. I just need to find enough clips.
@Generation3dX Alright it works fine using "she slowly and controlled pulls down her tanga to her knees, she is doing this with a natural motion." Atleast 10 steps needed.
@teeay25225 cool share a result when you get a good one.
Damn this is amazing, thanks man