I'd appreciate some yellow Buzz if you can spare it - I'm trying to buy a 12k model here and anything helps toward that goal :)
************
As the name implies - This is detailed Pussy for LTX 2.3 .
Trained on 5k synthetic image dataset. I generated a wide range of angles, ethnicities, lighting conditions and poses.
This is purely a detail lora to get proper pussy and anus detail. No motion videos have been trained on. Any motion gained is pure luck.
That being said - it plays well with other existing motion loras found here on Civit. I see potential for training simple motion LoRAs to accompany this LoRA as this will do the detail heavy lifting without conflicting motions.
All example videos should have workflow embedded - if they won't load - there should be a single image uploaded as well that would have the WF embedded. It's nothing fancy but it'll give you a start.
PROS :
Good pussy and anus detail, learnt a couple of erotic poses as well.
No face modification. Works well T2V and i2v.
Even does realistic breasts.
CONS:
Too many breasts and nipples! through shirts! I realized too late that my dataset had too many full nudes in it and thus it very rapidly learnt to draw nipples and exposed breasts. By that time I was already too deep into training and didnt want to fix it.
I tried compensating by merging the 'best-breasts' lora found here on civit at a -0.1 value. It didnt fix the issue but it helped a little bit.
Avoid prompting thin clothing, prompt thick jackets and wool sweaters if you want to keep the ta-ta's hidden.Only does fully shaven pussy right now. Majority of the dataset is clean like that. Might be able to blend it with some other hairy loras here for more fluff if you need.
Training details for those curious:
Trained with Musubi trainer - LR 0.00006 for 71,000 steps.
Some decent labia detail had emerged around 30k already but I wanted to push it longer so see if it would pick up on some of the subtler labia spreading details in the dataset.
At 30k I dropped the LR to 0.000045 and let it run . The LR was so low I could afford to let it run for a while without overfitting.
Images train much much faster than full videos and use much less VRAM. I could fit this into 24gb if I wanted to.
I'd suggest always doing a first run of image based training to establish details and then take over with your video motion dataset. This way you build a solid base understanding of the concept quickly allowing the video to come in and finish it off.
Description
FAQ
Comments (46)
does it work as breast lora?
almost too well definitely yes.
It's probably the best nudity lora in my opinion
@FantasticFever i personally find the nipples too ai perfect. dataset wasnt well made in that way.
@QualityControl I don't mind, gotta leave CV knowing it aint real!
mix with macromastia lora for good results
Doing a 2 woman snarkle now, just in low pass gen preview I already know this lora is a banger, nice fking job!
Thanks! LTX needed it
dare i ask what a snarkle is
@markdalias they went ocean diving without gear or snarkles (the thing that sticks out the water to breath with) They only had "toe fins"
@markdalias I think he meant 'snorkel"
Yes, I gravely misspelled a word and apologize to the brethren.
@FantasticFever yo I just got worried I missed out on some new goon trend
Any plans on tackling breasts with LTX 2.3 at some point? Right now, that is one major thing missing from LTX. Yes, there is one available, but it was trained on a very sparse set and does not seem to work at all. If you check some of my videos, for example this one: https://civitai.com/images/119112710 I have few short ones where a furry simply moves for example a towel out of the way, to reveal their chest. Wan has a multitude of Loras where you can specify the size and it stays consistent and what is revealed just makes sense. LTX does not quite seem to understand that, so when a piece of clothing is moved aside like this, some type of body horror or exaggerated breast size comes out that make no sense. I'm not looking to undress anyone, I'm looking to just create simple, sensual videos like the one above for my own Loras and whatnot.
i would say breast details are probably one of the easier things to train yeah. LTX knows that women have nipples, it just doesn't have the accurate detail - LTX had NO knowledge of pussy detail so it was a bigger challenge.
breast details should be simpler. I'll give it a go sometime
@QualityControl I suppose I should clarify. The issue isn't getting breasts and nipples. It's mainly controlling size, appearance and such that is the problem. But you've seen the video and probably understand what I mean, even WAN needed a dedicated breast size/small breasts Lora in order for medium to large honkers to not pop out from behind that towel. Want itself seems to default mostly to an average to large size of you don't use Loras or specify anything.
Either way. I look forward to it!
@OhNoMyChickens yeah. I understand what you mean. it comes down to a properly captioned dataset when training that.
currently most datasets are just captioned "boobs" and thus they congeal into one singular look like what my model does here.
I mentioned it's easier to train than pussy because you have to fight ltx much less for it.
@QualityControl But how would you make one that covers both the real deal and the furry kind that often has weird colors and such? I guess the model itself will interpret what it sees and interpolate to make things look somewhat alright. At least for my use, this is all that's missing and then LTX is on par with Wan in terms of the things I gen.
@OhNoMyChickens This is why ppl love the ability of Qwen and Klein to keep such consistency in image editing. cause you can take one image, then rotate the subject or whatever, and use all the outputs as training information. You have many angles/actions of each thing.
This is the best LORA I've seen, it works perfectly with i2v, undressing, breasts, vagina, it does everything, thank you, looking forward to the update!
This is a very nice lora for just images bro nicely done. Planning on redoing the original puss-daddy crap but a full rebrand. i like to start fresh , formatted, deleted old accounts, 400 video 1024 res 6 second Wan 2.2 dataset coming in a week or 2, taken 36 hours just to make the videos still got 72 to go ._.
https://ibb.co/8D2JfZJ9
image from dataset
@Brojakhoeman cool man. Hope it goes well. I checked the last dataset and I can see why you had trouble. waaay too many EXTREME closeups. Stay away from datasets from in-the-crack , it's too close and distorted.
@QualityControl yeah it was a little all over the place. Problem is with video I couldn't do high enough definition for medium shots to have any detail , I might also mix it with 1280 pixel images but I'll see. Gonna take so long already 😂
Completely re written the caption tool as well, from the ground up so I hope the captions will be better
I hate how 512 Res is 8s /it but 768 is 20 seconds 😩
As every video source is from wan I'll also caption each video with the caption used to make the video to begin with. Probably help people get better results with less input
@Brojakhoeman I'm gonna test my own fingering motion lora in the coming weeks. will train for 20k steps on ONLY images first to establish a baseline detail understanding and then pivot to only motion. hopefully that shortcuts to getting great quality without overtraining motion.
@QualityControl that sounds like a task for sure, i'd be tempted to say build it off your current pussy lora so it has a solid baseline and then focus on the motion with images at each stage - could be fun interesting to do from fresh tho for sure
@Brojakhoeman if I was to do that I want to start fresh because the current lora is too aggressive with the nipples. Next round I'd train it with more nipples covered so that there isnt such an extreme bias towards boobs out.
careful on doing this with your own too.
ALSO -
I noticed your pus-daddy lora had a HEAVY face influence and ethnicity bias. Made everyone look the exact same even at 0.2 power.
I generated by dataset with a large range of ethnicity and face wildcards to mitigate the 'same-face' issue and it paid off because it seems to have stayed flexible. Make sure you get dark skin tones in there.
@QualityControl noted - this time around there's no faces 99% neck down content, - i managed to get multiple skin tones but for some reason black women here on wan, straight it refused to gimmie black pussy lol meh. Just finally finished my new caption tool - now it uses claude locally to caption 10 images from each video, much smarter and faster then standard qwen...
First off, just wanted to say that I love this lora. It works really well to generate a realistic puss. It's unfortunate that it has a strong influence on breasts in i2v though. They're pretty tits, but it overrides the reference image breasts strongly. You can watch them change over the first few seconds of any video. Lowering the strength helps a bit, but then the puss doesn't look as good. It also has a mild influence on the face. I'd love for this lora to be purely Puss focused. Anyway, nice work!
yeah. I paid 0 attention to breast training in this run. it happened by pure coincidence and it's too strong I agree. I hate that t2v gens have nipples poking through.
On a next run I plan to address that with a better dataset
Thanks for the hard work. I'm looking forward to the next revision.
Using both of these enhancers at .6 helps make the skin more realistic, making pusssynth great for t2v at .9-1.0, just wish men didn't get boobs.
https://civitai.com/models/2535622/ltx-23-enhancers?modelVersionId=2849706
My example video. https://civitai.com/images/127293778
try a combo with my Amateur Hour lora as well. it will do similar, and it was also trained on some nudity so it'll add more realism there
@QualityControl why yes, I think I will
When trying download writing error getting file. What wrong with that?
It's not the model's fault, it's the main site, the backend, the cdn and the site routing that have been broken for at least 60 hours if not more
probably the most impressive female genitals area and anus quality in LTX2.x out there (tried myself a backside lora it was trained on videos only... had hard time going over 768 resolution on my 24 gb card, took ages and result was nowhere near this quality) I don't know how to use musubi trainer on windows though, so I can't (using Ostris A.I toolkit, which apparently don't accept images training on LTX 2.3)
musubi is annoying to use yeah but any llm can help you get going if you show it the github and ask it to hold your hand to set up the scripts.
great i want to try and train one myself. hows does one caption a dataset like this to only train the P---y?
caption it any way you want. I captioned it fully and made sure the vlm mentioned genital details and position so that if someone prompts for those details then the data exists for it.
dataset variation takes care of the risk of likeness burn in.
my lora here fails at keeping the titties contained which is a flaw in my dataset, I have too many topless images in it, so it learnt to always show nipples. make sure your dataset is varied enough in different states of dress. even include some clothed data.
Works amazingly well! Works awesome for close up shots. Movement is very realistic.
I'm impressed you got movement because there's no motion trained. I guess a large dataset with so many varied poses just acts as 'motion' in a way. Glad you like it
@QualityControl Yes, I'm surprised too. Movement is actually very natural.
Is there a trigger word for this? and a preferred Image to image workflow?
Amazing loRA thanks :)!
Keep training G! And don't think about skipping even a day! We all appreciate your glorious training.
Details
Files
SynthPussy_01_rank32.safetensors
Mirrors
SynthPussy_01_rank32.safetensors
2509189_Synth_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
pus.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
pssy.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
2509189_Synth_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
LTX2_Pussy.safetensors
SynthPussy_01_rank32.safetensors
Synth Pussy - LTX2.3.safetensors
LTX2_Pussy.safetensors
SynthPussy_01_rank32.safetensors
Synth Pussy - LTX2.3.safetensors
synth-pussy-ltx-2.3.safetensors
Synth Pussy - LTX2.3.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
SynthPussy_01_rank32.safetensors
Synth Pussy - LTX2.3.safetensors
Synth Pussy - LTX2.3.safetensors