NEW MiniMax H3 version 1.0 !! - all samples are pure t2v, not i2v or ref mode

This took a lot of experimenting. Rebuilt the entire dataset.
Results came out quite nicely.

this was trained on about 500ish synthetically generated images on a 'expected' look of pussy and labia detail. real vaginas are so wildly varied it's tough to build a dataset which a model can easily learn - so I made a Synth dataset I more accurately control and thus it learnt better.

All images in dataset generated by Krea2.

Most if the dataset is pussy shots front and rear. I included closed legs standing as well as lying down spread legs for positioning variation.

As a bonus I included some fingering and vulva spreading and object insertion images. This helps teach the model what a pussy looks like under those conditions and I think it worked out very well.

0 videos were used in the training yet you can prompt for some actions like fingering and spreading and they just work.

Dildo actions will still need additional lora help, but you are able to promt objects inserted and the pussy will deform mostly correctly.

Enjoy! Go make stuff and post it here.


************
Old post - 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.