Experimental lora for a POV gay missionary sex scene. The position works very well, but motion is not very well trained. Use in combination with the straight missionary lora (https://civarchive.com/models/1087616/hunyuanvideo-pov-missionary?modelVersionId=1241027) at about 0.4-0.6 weights to help guide the motion. Describing the motion well in the prompt can also help (for example, another man at the bottom of the screen penetrates him deeply, with a steady rhythm, his penis sliding in and out/ hips thrusting forward and backwards/ fast repeated or looping motion etc.). This is definitely something to improve in the future.
Some helpful tips for the prompt:
Describe the camera position or how it holds steady if the scene shifts too much.
You can prompt for the position of the bottom's hands (behind head, on table, on own thighs etc.), but it's not always reliable.
Description
Trained on 27 videos and 8 HD images.
FAQ
Comments (12)
I love you so much, there's so little gay video loras D: Will definitely be testing this out and hope to see more.
perfect
looking forward to using this one! :D
It had to be done. The ai porn revolution must continue to.....flow.
This is soooo good. Can you please help us with how you trained this? I know motion is an issue, but it's still so good.
Hey thanks for the kind words! I’m not an expert or anything. I basically collected videos of similar poses, with diversity in the bottoms (ethnicity, age, body hair, etc) and a few HD images. I processed the videos so that they are about 3-6s long, 24 fps, and extracted some frames from them. The frames and images were then captioned with Claude 3.5 Sonnet. I emphasised having the trigger word POV gay missionary scene of … at the start of each caption. I realised now I forgot to explicitly caption the motion of penetration so that probably contributes to the motion issue. The images, videos and captions were then uploaded to a cloud-based service (I used MimicPC, but I guess it should be the same on Colab/Vast/Runpod/local) and run for 50 epochs total, 10 repeats, 2e-5 learning rate, saved every 5 epochs and I pick the best one. I find typically for my Loras epoch 40-50 work best (there are tools that plot loss to pick the best one, but I’m not familiar with them). Hope this helps! It can get quite expensive though so I’m still trying to figure out the most economical cloud option since I don’t have a powerful PC.
any sample prompt?
Can you upload the Workflow. This is fantastic!
Thank you
We need gay content like this, thank you! <3 Too much puss on this website LOL
amazing
Notes on how to train this type of content,
by OP Petrilover :
- I basically collected videos of similar poses, with diversity in the bottoms (ethnicity, age, body hair, etc) and a few HD images.
- I processed the videos so that they are about 3-6s long, 24 fps, and extracted some frames from them. The frames and images were then captioned with Claude 3.5 Sonnet. I emphasised having the trigger word POV gay missionary scene of … at the start of each caption. I realised now I forgot to explicitly caption the motion of penetration so that probably contributes to the motion issue.
- The images, videos and captions were then uploaded to a cloud-based service (I used MimicPC, but I guess it should be the same on Colab/Vast/Runpod/local) and run for 50 epochs total, 10 repeats, 2e-5 learning rate, saved every 5 epochs and I pick the best one. I find typically for my Loras epoch 40-50 work best (there are tools that plot loss to pick the best one, but I’m not familiar with them).
- It can get quite expensive though so I’m still trying to figure out the most economical cloud option since I don’t have a powerful PC.
just for whoever might come across this in the future, this great info was buried in a comment