READ PLEASE
I wasn't planning on uploading this model so early because it needs more training, but perhaps you can help me with your feedback. The model is already quite capable with the "pussy", but it struggles to generate the anus in some cases, so you need to try different weights and seeds (it needs more training). Perhaps you can find better combinations or weights that work well; your feedback would be helpful.
Please keep in mind that this is the first time I've trained a LoRa model with my humble hardware, and I'll probably make many mistakes. I'm only certain of one thing: the dataset is very well captioned because I did it manually (yes, I described 50 images in natural language and in a variety of ways without using AI).
Use pussy and anus instead of vagina or asshole
Technical Specifications:
Training Software: Fizgig (H3 is still experimental here)
Epochs:
Done: 35 (1750 steps)
Remaining: 25 (1250 steps)
Training size: 0.25MP with 0.25 Detail focus (I don't even know how this impacts, but I'm testing; 0.5 was taking too much time). Trigger word: pussy4nus.
Recommended strength: From what little I've been able to test so far, the best values are between 0.7 and 0.9; more strength generates artifacts, noise, and aberrations, while less strength results in a very poorly formed pussy that looks like a cut. Anyway, please help me find the sweet spot.
I used poses from behind, from below, frontal, closeups and even on side, so keep it in mind, this model will handle every existing pose. (i hope so).
IMPORTANT SETTINGS USED:
Workflow used is made by @joeygambino cheers to him!
Turbo Lora at 1.3-1.5 STR - Sage Attention - 8 steps - Euler/Beta - 0.35 and 0.4 MP
Useful prompts:
the woman is exposing her clean anus and pussy, the anus is on top and the pussy is beneath the anus, her two hands are holding her buttocksfair skin and big buttocks exposing her nude anus and pussy, her anus is on top and her pussy is beneath the anusKNOWN LIMITATIONS:
(needs more training): The anus sometimes doesn't generate; you have to help the model with the prompt. Below I'll leave some phrases I used in the screenshots to help the model recognize the anus's position.
I didn't test T2V, so it's up to you guys, i can't test because my GPU is already busy training lmao.
I didn't tried sex scenes with penetration yet, so feel free to test the pandora box.
Description
A simple preview, read the descripcion PLEASE
FAQ
Comments (24)
please READ the description!! 🫶
bro pls next version
this is why we can't have nice things.. everyone is like gimme gimme. make the next version yourself- leech.
@axicec bro pls
@yuduz367 bro imagine asking for next version the moment a version drops.. seek help. touch grass
@axicec I don't know what's wrong with you, but you should isolate yourself from society. You're weird.
I made this https://civitai.red/models/2835594/h3-vagina?modelVersionId=3200540 and mine just kinda broke the more I trained, I think because the model is distilled and not a base model. But I used ai-toolkit. Do you think fizgig handels this better? Or maybe your captions/dataset is just better, idk. But good job :D
Yeah i tested it and worked but i think it needs more training (? maybe. to be honest i'm making this because i want to have pussy and anus+spreads of my taste, i never used AI Toolkit, i saw there was this Fizgig with a cleaner UI and i liked, you can use int8 pruned to train + nvfp4 encoder. works fast and is easy to use but doesn't supports videos, at least not for now as still experimental. and about the captions i had to make it handled because honestly AI sucks managing different angles and details so i did it manually and using multiple poses and multiple different phrases
I did a split around 30 pussy and 20 pussy+anus, including medium, close, closeup and extreme closeup shots
I think yours is really nice too herpderp! :D I'm in the middle of comparing and contrasting yours with the one here, seeing which ones give me better results. It's still early days for this model, so having a vag lora this early is amazing enough as it is thanks to you guys.
@dirkzen i'm posting a new version tomorrow with x2 of epochs
Is this for the ref model or the i2v model? I think most people are liking the ref model better. It can even kind of do the i2v ff lf stuff too. if prompted for it well.
Honestly idk, i trained on i2v, but maybe works on ref model, can you try for me? i don't have the ref model downloaded and can't test cuz i'm training
i prefer the i2v actually, i can't prompt the ref model to continue exactly from one frame, it always comes up differently so i can't stitch videos together.
bro pls previous version
I tried the second prompt without the lora and it did the same thing and the result was better.
post the video so i can see what to improve
@simonishere Keep training bro all these loras are undertrained like crazy it's gonna take a while. Keep going until it starts looking like it. If it becomes corrupted somehow then settings or trainer needs changes.
@simonishere I didn't play with it too much the main thing is the face changed with the lora.
@psvr2_wins feels like an expectation if the dataset includes a lot of faces, it probably learns those incredibly easily. Also, fizgig I think only takes images, so that'll impact the outcome as well
thank you for making it a cleft of venus pussy and not sleeve of wizard
whats a good weight number?
0.8, anyways this still training, maybe i update in the morning
So far the only weight that worked for me was 0.5, but im using first frame and last frame, anything above 0.5 and the video starts out with a very close up shot of the vagina/ass.