Features
Compatible with both T2V & I2V Wan2.2 Models
Trained on simple word tags: vagina, anus, ass
Tips
Look like strength between 0.5 to 1 is good. I normally use 1 for I2V but others are mentioning that 0.5 works, too.
My Wan2.2 Cover Videos
Workflow is T2I SDXL/Flux/Pony -> Image Upscaler -> Wan2.2 I2V -> Upscaler -> Add FPS
https://civarchive.com/models/1916073 (text2img & img2vid). Credit DonutsDelivery & Daxamur
Training
Trained on Wan2.2 T2V base model. 50 high quality images and 50 videos that were handpicked and farmed raised to perfection using musubi-tuner. This LoRA works with both I2V and T2V.
Limitations
The dataset contained videos with hand movement like rubbing legs or rubbing breasts. Some of the videos also had slow hand motions. You may need to prompt to remove or proactively address by saying where the hands go, e.g. hands stay still, hands are still, hands rub her legs.
Pussys are extreme complex. No two pussys are the same so the model struggled at some points to fully generalize the pussy's complexity. More training may be required in the future.
Description
FAQ
Comments (36)
Oh my god, a Loras 1GB? Is it really necessary to be that large? Will this take up too much storage space?
@wange999 its two LoRAs. I had to increase the models parameters to better handle La Pussy. I may do more testing to reduce the size but need to make sure it doesn't reduce the output quality.
@Kierkegaard420 Unfortunately, my storage space is running out. Models that take up too much space will have to wait until I upgrade my storage. However, I still want to express my gratitude to you and the creators in the community. All of you are worthy of respect. Thank you!
@wange999 No worries - thank you for the feedback. I'm working on training another model right now but hopefully will run some tests to compress this model a smaller size model.
@wange999 a very kind member of the community has just messaged me. they've "compressed" the lora model to roughly 96% accuracy to the original: https://huggingface.co/JustAnotherCibrarian/send/tree/main
@Kierkegaard420 Thank you!
1GB is a lot? 🤣
Guys we need to gofundme for this man so he can store 1gb pussy lora
I'll take a gigabyte for 100% accuracy over a compressed 94%
anus? ass? its not just pussy?
@CharlieBrown0115 Yes, for some reason the model seemed to train better when it was trained on the word "vagina" vs "pussy". You can try generating using the word "pussy" and let me know. The training may have associated the images/video to the text pair, "pussy". However, all my tagging used "vagina".
Anus and ass are there because you can have images/video that don't contain an ass or anus. It allows the user to be more specific when generating prompts. These tags normally exist in this type of training, I'm just highlighting them for users to know about.
@Kierkegaard420 nice ! so , is not only about pussy cats! its the whole combo!!! cheers!
I recommend using this with 0.5 on high / low and mixing it with the Femaled-Vaginus Lora with strength 1 on high / low. Someone already uploaded an example ;)
Just a personal opinion after some tests ofc!
@Lora_Addict Thank you for the input and suggestion!
还是会改变人物的面部，而且会让丝袜消失
16 000 steps man, how many times with which graphic card ? I try diffusion-pipe and ostiris tool kit, and 5000 steps it takes on entire night even with b200 ,, wonderful model btw ^^
@Meandre66 It was 16k per model on a 5090.
@Kierkegaard420 How long did that take in total and how much vram was used?
@vigilence Roughly 40 hours total for both models and I maxed out the VRAM.
Will Lora continue to train the content related to male penises? I found that the original wan2.2 i2v doesn't understand the male reproductive organs. The generated content is mostly broken.
@wange999 I'm training another LoRA right now but I already have a list of 'fixes' for version 2 planned. I'll add this to the list.
it changes breast type i2v
@macehih402333 What?
Maybe he means it makes smaller breasts bigger.
great lora but it changes faces! remove faces when training...
@sirbazzrick How did you split the High and Low noise for your steps?
I'm actually amazed that it took this long for someone to make this
This would make a great inpaint lora, if used with a mask, otherwise due to the inclusion of faces in the training images, it changes the output faces way too much.
So you're saying that the chocha comes with a face?
It changes faces.
Can't wait for the latest version😋
Could anybody please upload this model to TensorArt?
I spent 3 days finding a lora, that can creat good female anatomy if its not on the base image, i tried like 5 or 10 very popular lora, their animation, and face respect are good, but damn, non of them can creat REAL one hole or vagina.. i can finaly sleep now, Thanks , and good job.
So on the hole, what exactly are you saying?
Thank you for this. I'm still trying it out, but so far it seems to work beautifully. Suggestion for another lora: small nipples. Or something to define what kind of pair pops out. Every time I prompt for clothing removal, default, quite large, kind of nipples appear.
Sadly this lora changes the face way too much. Only thing you can do is using it at like 0.3-0.4 and mix it with other V loras.