Trained on 167 lesbian cunnilingus videos [[544,960],[960,544],[720,720]] x 238 frames @ 24 fps (about 10 second clips). Both faces always visible
I tried to label innie vs outie pussies and orgasms
----------------------------------------------------------------------------------------
V2 is retrained from a checkpoint in the middle, with more of an orgasm focus, may be less flexible, details seem less clear
---------------------------------------------------------------------------------------
for V3 I tried training from an earlier checkpoint and increasing the resolution (reducing to 6 seconds) on fewer, select clips, and halving the learning rate
tbh I wasn't going to release v3, cuz it just doesn't seem to do what I want it to but I said I would
the tongue does seem to move more, and the details seem to be better than v2 - there is much less of an orgasm focus
sample generations were made with speed up lora minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors, I got it here on civit
I don't know if I will be updating this again, this may be a job for people who are better at this
Description
started from earlier checkpoint, did some retraining on a focused dataset
FAQ
Comments (18)
is there going to be a man going down on a woman version of this?
not from me
What would be the fun in that? LOL!
I'm not positive but I think the shaking in your samples is from your LR being too high during training. If you tried like a 5e-5 might slow down the shaking
perhaps, my default lr is 1e-4
I love H3 but jesus christ does it hate anything to do with tongues. I've been fighting with my kissing lora since the day the model launched - not trying to be negative but it seems like your lora struggles with it too.
It's almost impossible to get it learn intricate tongue anatomy/movement.
yeah it's pretty hit or miss,
describing the motion at inference time can help some, but I feel like WAN just kinda worked
I've had a LOT of trouble with details, apparently it compresses the latents heavily, you need to feed it the highest resolutions you can find, add high res images if VRAM doesn't allow videos to add the detail...
@az420 I might try to tune it with a later checkpoint and some higher res clips then when I get the time
@anonymist I've tried adding 4k anatomical tongue reference images into my dataset and it didn't seem to make much difference. Let me know if you find out anything, good tongue action is pretty much the only thing this model is missing atp.
Is there some 69 clips in the dataset ? Getting this sex position with vanilla H3 is quite difficult.
Great 2.0 update!! Very hot!
works like magic! great work
This works pretty darn well. You've noted the short comings, but v2 produces fun results.
Edit: I will note this does hit quality in a way that's unfortunate. The content is top notch, but I wish it wasn't so prompt to introducing distortion and softening details.
yeah I think I was too hasty with the upload, I am working on a new version, hopefully it turns out better
@anonymist The action is so good. You're on the right track. Try adding a bunch of anatomy photos.