This model is fully compatible with S2V It can bounce in rhythm
This is an i2V but i try some t2v seems to works too, character may move her body like she have a vibrator in her pantie without the vibrator sounds, just the body reactions. she moves her hips up and down with a "ahegao" face style , audio produces clapping and moaning and inhale exhale noises, sometimes good , sometimes bad, if you can use negative( not on distilled so cfg>1) put "music" in order to prevent some random sounds/noises
The videos i will publish may talk a lot more than this description
It works fine on distilled Quant models, 10s 24 fps is fine but 10s 50 fps is better 720p or 1080p
In fact 50 fps is better because the body may move too much to be interpreted in 24 fps (lol)
Experimental training I will publish 3 lora because all of htem is a bit different
The first one is the most aggressive/wtf 3750 steps
The second one is less aggressive more erotic but seems to produce random waters on some seeds 5250 steps
)The final one is 8500 steps it may have more moves variety and flexibility but not really sure
Concerning the S2V possibility, you may prefer base model , it is better for that ( but long to process :/ ) , the lora is less efficient with the base model but more realistic.
Sorry for the files name, wasn't really sure of the name at the start, i know some use tool to download and that a mess with random names
I didn't test too much on non distilled model, because too long on my local machine and i can't have more than 540p 24 fps with 30 min of delay ,
Trained with old grok Ai imagine video more than 100 vids 145 frames with audio
Description
FAQ
Comments (2)
Can you share with other creators what you did to make this? The output quality is INSANELY good and higher visually quality than any other lora I've ever used. Like it's good as just a general quality control lora. So many other loras here make blurry outputs that degrade a lot from the starting frame.
Actually nothing really special, a dataset with at least 200 video 6sec 24 fps made with grok (when it was not so censored) , trained with ai toolkit setting by default for ltx2 just change the number of frames 24x6=144 +1 =145 frames, i just put the same prompt for all videos as it was same set of moves for all videos.