About this Lora:
This Lora will create videos of a side view of woman and a man having sex in Doggystyle sex position.
About Buzz:
Really appreciate some, as it is difficult for me to get my hands on crypton due to government relulation. So if you like what you get for free, please consider helping me out, thanks :-)
Testing:
• Actually really reliable at strenghts from 0.4 - 1.0
• It should not really influence the chars faces as I intentionally cut out the heads in the database.
• Works great at 0.45 strength and fixes issues like style bleed or cropped videos.
• best used at 0.5 - 0.7 (t2va / ref2va) and 0.8 - 1.0 (fl2va)
Where it shines:
• Very reliable motion
• Adjustable breast size
• no sound interference
• decent usability with ref2va
• works great with different angles and positions (standing / on all fours / and more)
What it struggles with:
• It cannot know what a pussy looks like, as the dataset did not provide well enough. Maybe I retrain it with some more anatomical detail.
• It does not really help to thow additional loras at it to improve pussy details (but please try)
• Sometimes the Penis does not really slide inside in a clean manner (additional loras can help here!)
Promt example: [just use natural language]
Male and female have sex in doggystyle position on [Setting]. The female is bent over. The male is thrusting his hips back and forth rapidly, while pounding the female's ass. The stiff thin Penis is penetrating the pussy vulva. The female breasts are swinging back and forth by the force of the pushing male.
Description
FAQ
Comments (17)
Holy crap is this lora amazing. Start with a ref image and watch the magic unfold. Thank you for making this.
Thanks for making this. You mention that it does not have enough anatomy data, but does it really matter? MysticXXX knows almost every position. So i would use this lora as a motion helper only anyways.
I agree, i probably does not matter much, especially when using i2v (i.e. fl2av)
Fantastic work!!
The movement is good but this completely destroys the Audio
sorry to hear that. i did not exerience that this much. maybe because i use it at lower strength. i guess it happens with ref2va? maybe you can try use advanced lora loader nodes and disable audio for this speciffic lora.
@Adaptalab0r Yeah, I have to back this up. Results are really good but audio in ref2v is completely unusable.
Foley is file. It's specifically any dialogue that comes out distorted like it's super compressed compared to the same gens / workflow without this lora.
I was using it at 0.7. Let me try it at 0.5 to see if that helps.
@Adaptalab0r correct it's happening with ref2va. Perhaps look to use some normalization audio in the training.
@thomas852002887 Do you mean normalized audio in the dataset videos or do you mean speciffic settings in ai-toolkit? I used the on site civitAI trainer. there might be some switches to look after, but i doubt that there was one speciffically for audio.
Very good LoRA, can you share some general details how did you train it?
twenty-two 24fps videos 640x480 and 800x600 videos. some heads were on it. sowme were cropped out, depending on perspective. I wrote the captures manually just using the same text blocks depending on whether the penetration actually was visible. the capture was simple and concentrated on the pose and some elements of the setting i.e. on a couch, living room interior setting.
i trained it on civitai 10000 steps but after 2000 steps the motion was already decent enough.
@Adaptalab0r thanks!