Example prompt using ImpactPreprocessorWildcard ComfyUI node:
A feminine transgender woman with {long|short} {blonde|brunette|red|black|purple} hair {tied in a ponytail|tied with a bow|untied} is riding {a man|a transgender woman's} erect penis in a {dimly lit|well-lit|brightly lit|dark|sunny} {bedroom|living room|gym|outdoor location|public location}. She is fully nude and is bouncing up and down. She has a curvy feminine body with natural feminine breasts and a naturally feminine face. Her full body and face are in frame and her penis is in full focus. She has a beautifull feminine face with an orgasmic expression, looking at the viewer, maintaining eye contact with the viewer, Realistic, ultra detailed, best quality, ultra high resolution
Description:
Version 1.1 was only trained on realistic videos to get a more focused output. Some of the training data contained cumshots, so this model may also do that. If you'd like animated output, use v1.0.
Training:
I trained on an A6000 using 20 2 second videos (24 FPS, at least 48 frames each). Training 50 epochs took around 24 hours.
Description
FAQ
Comments (22)
I wished that prompting would be easier than writing a whole Harry Potter Book for a 4 second video :D
Your prompt shouldn't have to be as long as the example here to produce good results.
Can confirm, works with even a single short sentence.
But i see that most prompts are so long for the video generator :D
I'm quite happy someone's finally making some futa/trans loras, could you make a solo masturbation one, by any chance? (I wish civit was able to train these so I could make character loras for Hunyuan, but...)
Yeah getting more futa loras is great. One thing I noticed is that you can get base HunYuan to do futa pretty consistently if you prompt it right. Also you can get most non-futa loras to do futa pretty well too, depending on how flexible/overtrained it is.
@Degenerator123 huh, really? I fought with it a bit and could never get anything that looked right.
I could probably make a solo masturbation model pretty easily. Should be plenty of example videos I could use for that, even more than this model.
@pikenrover You should do a cum on self one. Not for me its for my friend or whatever
Def need a solo futa model, preferably with cumshots.
Did you use vid2vid? I'm not getting anything close to this smooth of an animation with the sample prompts you provided.
No, I used txt2vid using Hunyuan Video. I don't recall the tutorial I followed but there are plenty out there.
@pikenrover gotcha. thanks!
11/10, the motion is incredible!
awesome, please train more loras! Can you please do futa dom + smaller female? or futa/trans + onahole
Fantastic lora! Takes only a bit of playing with the prompt to get futa+futa consistently. One thing I noticed in photorealistic gens is that the girl riding often gets a sort of navel/bellybutton piercing. Not a complaint or anything, just something I noticed.
Thanks for the info! Perhaps the next version can account for body piercings.
@pikenrover It's not a big deal and it doesn't happen every time, I assume one or more of the samples in your dataset had someone with a navel piercing
The examples look amazing, but I can't make it work correctly. For some reason it ignores the prompt (I followed your guide) and the futa is always with the back to the camera. Always. Even if I change my prompt, it doesn't matter what I write, she is always in non-reverse cowgirl position. May be there is something wrong with my settings, I don't know, but other loras work well.
Any clue what may be happening?
Thanks for sharing this, in any case. :)
Yeah for me I just bruteforce it with different seeds and eventually get some good ones. Don't know how you long you wait for generations but generating smaller resolutions to find a good seed and then upscaling that is what I do
This is an issue I also noticed when testing this model. I'll look into addressing this when I find time to fine-tune this model. As @Akiro69 said, I also did this and brute forced generations until I got what I like. Better prompting might help. Shorter, more concise prompts that may include explicit things like "She is facing the viewer" and "her butt is not visible in the frame" could possibly work. I haven't tested this though.
@pikenrover I'll experiment more with prompting too. If it's anything like how LLMs were for a long time, positive prompts will work better than negative ones, for a example, rather than "the woman is not wearing clothes", instead saying "the woman is naked", "the woman does not have hair" vs "the woman is bald" and etc. It's possible Hunyuan works differently though.