Didn't turn out the best but spent like $6 training it so might as well release it, trained on 7 x 270x480 videos, play around with the strength. My example images used .90 lora strength and 9 cfg.
Only seems to work well with pink dildo if your using fast hunyuan lora+model
Example prompt :
dildo machine, a view of a woman leaning back with her legs spread having sex with a dildo machine. She is wearing a pulled-up white top with her breasts showing and has long brown hair. She has a tattoo on her chest. She is on a white sofa. She is moaning and looking left. A machine is thrusting the pink dildo back and forth inside her pussy slowly at the bottom of the screen. The camera is holding steady. The scene is in a living room.adding something the following can help if your cfg is high enough
dildo penetrates into her vaginaDescription
FAQ
Comments (8)
"Only seems to work with pink dildo"
This could be an issue with captioning. Did you caption it "pink dildo" every time?
Could be, I used pink, light-pink & white once, the data was mostly pink but maybe I should have captioned it more generally rather than specifically since my dataset wasn't that broad
@onbe
"Generally speaking, you caption things that you want to exclude from the training."
I find this article helpful for captioning: https://civitai.com/articles/7203/captions-vs-no-captions-a-deep-dive-into-effects-on-flux-lora-training
But if the majority of the dataset contained pink dildos it still could have trained it in, if it was 'overtrained'. If you have them, I'd recommend testing a lower Epoch snapshot and see if you can get a different color out of it.
If you do decide to retrain at some point, adding in some additional colors would likely solve it.
@TheAIDoctor seems like its actually a compatibility issue with fast hunyuan + fast lora. Able to modify the color etc in normal hunyuan video. 🤔
@onbe I've had pretty bad experiences with the fast Hunyuan model so far, especially using Loras so I almost always stick with the BF16 version. But I know that's not always an option for people with less VRAM.
@TheAIDoctor I have 24gb vram but I don't like waiting more than 50 seconds xD My latest image used the gguf hunyuan video q5 model and it seems to work much better now. Will read your article though and see how I can improve prompting in the future. Ty
tattoos for everyone too! ;)
great idea though!
Would be great if it could also work from side in 16:9 aspect ratio.