Sorry no. Focuses 100% on females from various races and hentai in many different styles - from grayscale sketch all the way to full anime style. But the LoRA struggles immensely with drawing genitals from nothing so if you just use it as a motion assistant and you have a good starting point (image) you can still animate it properly even if both characters are male - I think.

Thank you for the kind words. Still its far from perfect. There are many concepts in my datasets that the LoRA simply refuses to learn and despite the fact I changed my datasets to enhance genital learning - it still struggles to learn both pussy and penis at the same time. I barely tested it myself so I do not know how good or bad it truly is - but from the little testing that I did I can say there are so many flaws with it if I started writing about them all this comment would turn into a LTX-style prompt

Yes I'm planning on making a new version with major changes but it will take me about a week - all things considered. Plus, I won't release it if somehow it turns worse than this. No promises but I'll try

Its been 9 days and I said it would only take a week so I feel like I should update on this.

I continued training on top of the released version and came to realize it contains heavy biases and is very overfit. Not possible/ideal to even attempt to fix that so I'm training again from scratch and this time I'm using a much lower learning rate. Which means it will take much longer than v1 did.

On top of that I'm gonna experiment with trying to train single-concepts loras on top of 'lora v2 - base' (once that's done) and merge them wisely into the base lora based on which blocks are affected more by each concept. All in all this is really gonna take some time and I can't even say with confidence it will be worth it - but I'll do my best.

This will be a bit long.

Ballpark of 300 images though they were mostly for teaching genitals - not sex positions. Though to be honest it doesn't seem like they were of much help. The model struggles to learn both penis and pussy at the same time and it didn't help that I was also teaching 'puffy pussy' and 'spreading pussy' in the mix.

For videos it was roughly 250 T2V videos with WAN-like captions then I duplicated those into a separate dataset and simplified their captions to be just mention 'motion/what happens', ex: 'having sex in the cowgirl position'. Those with short captions were trained exclusively in I2V mode while the others exclusively in T2V mode. For this to be possible I had to add a few lines of code into 3 files of the Musubi LTX-2 branch files.

I kept downloading more videos but I was lazy with the captions so I added the extra ones just to my I2V datasets.

At the end I still had ~250 T2V and ~400 I2V videos. 121 frames 24 fps - all of them.

Trained in multiple stages, first at LR=0.00008 gradient_accumulation_steps=1 for roughly 4 to 6 thousand steps initially using only [512,512] buckets for videos and both [512,512] and [1024,1024] for images. I ALWAYS train videos on FULL frames. I don't trust whatever Musubi is doing with other 'frame_extraction' modes.

I interrupted this run halfway to manually test it and found out that when used with the upscaler -> the upscaled video would look very bad. Kinda seemed to add 'noise'. Figured it was because I was not training the videos at higher res but with my 16GPU I'm kinda limited. So what I did was create 49 frames 24 fps clips out of all of my videos that could be used for [1024,1024], [1280,720] and [720,1280] buckets. Did this to both the T2V and I2V datasets and ofc I had to re-caption many of them because suddenly some parts of the captions no longer happened in the shorter clips. Added those to my training configuration at the mentioned buckets based on their aspect ratio and once again trained on full frames. This solved the upscaling problem.

At some point, I classified my videos in 'tiers' from 1 to 4 based on how visually pleasing they are to me and how good their animation is. I added extra separate datasets for tiers 2, 3 and 4 to serve as extra repeats for those 'better-looking' videos.

Also at some point, I wrote a custom scheduler that would multiply the base LR of a sample based on either the filename or dataset directory name. I2V datasets had higher multipliers and the files of tiers above 1 also had higher multipliers as their tier increases - something soft ofc - the max a multiplier could be was '2.0'. This was kinda like 'increasing repeats' for those files but by increasing their learning rate directly.

Finally, trained further with LR=0.00005 gradient_accumulation_steps=2 for an additional ~2000 steps (where a step here is actual 2 micro-steps cause Musubi counts a 'step' only when the weights updates - which with GAS=2 is every 2 actual steps). I wanted to train it for much longer but even this took me days on my 16Gb GPU.

Important: I used '--lora_target_preset v2v' which is mentioned in the musubi ltx-2 branch to be for 'video-to-video/IC LoRA'. Good thing I ignored that and went along with my guts because I'm pretty sure for NSFW the regular 't2v' mode wouldn't cut it. I know that from experimenting - during training breaks - with weights filtering and comparing other loras that people released on this site. My overall impression was that: every LoRA that contained the extra weights used in 'v2v' mode performed much better than those who did not.

That was even longer than I thought lol and I think I'm forgetting a bunch of things but this basically covers at least 80% of what I did.

Yeah its way more overfitted than I thought.

May sound crazy but I only noticed that yesterday after training for another whole day and testing the results. I released it without realizing - had I noticed these biases earlier I may not had enough confidence to release.

There is very heavy cowgirl and breast bias in its current state - prompting for 'small breasts/flat chest' in T2V mode will still give you at least medium sized ones or it will be completely ignored.

Now to make a new version I'll probably have to either start over from scratch with my newly modified and more balanced datasets or use a very early checkpoint -.-