I used Musubi Tuner because I only have a 4060ti 16gb gpu and 64gb ram. It has low vram settings where you can set the text encoder and model to fp8 to reduce vram too. It uses deepspeed with accelerate so I also went into the accelerate config and set deepspeed to it's ZeRO Stage 2 offload to further offload more stuff to the cpu from my vram. It gives me enough space with the 36 single blocks offloaded to train video clips on my gpu.

This lora was trained with 13 videos cut to 3 second pieces, and 10 images because I found it had the motion but lost the pose frequently without the images, like she would be bent over moving and he would be at her side and other weird issues.

All my videos came from redgifs and realbooru, and I had chatgpt write a python script to resize all the videos to around 384x214 while preserving their original aspect ratios so no warping. The different sizes is why the pose angle changes if you change the dimensions.

Musubi hadn't implemented setting steps per epoch yet, so setting frame buckets to 1, 30, and 75(3 seconds at 24.97 f/s) gave me 43 steps per epoch. I ran it for 85 epochs and that high was super overbaked so I moved down and epoch 60 was the one I liked the best as I tested different prompts and dimensions and weights.

@midiaplaay I just use the custom python scripts because I can use them for whatever. I used chatgpt to take all the videos I had gathered which some were a couple minutes long and split the videos into 3 second pieces. Then I went through and grabbed pieces I felt worked the best for what I wanted. Then I had another script create text files with the same names as all the clips. Then I had chatgpt write a script to reduce all the clips sizes to where the longest side couldn't pass 384 but the clips kept their aspect ratio. I tried using Taggui with multiple different vision models and none of them were great. They'd describe the backgrounds good, and that was mostly it so I had to manually caption everything.

For my images I resized the same way and had a script make a text file with the same name as each file with the word doggystyle in it and nothing else. The images came from thr clips from my first lora and got a script to pull a random frame from the 10 videos for that.

I pretty much automated nearly every step with custom python from chatgpt.

Posted the fix. I'll correct the main one in the next couple days

@midiaplaay all I did was manually describe the video in detail making sure to include doggystyle in it. Described body, skin tones, location, how they're positioned including for the man. Also described the motion because hunyuan already knows doggystyle but not very well. That's for the videos. The images I used to help force the pose just had doggystyle for the caption. I made a lora without the Images and it would frequently mess up the positioning with the 2 people but have them both moving correctly so I'd get stuff where the guy was thrusting into her armpit while she bounced like he was behind her and other funny bad tries.

I have no idea I could have done better differently.

@bleh0160 

new kink unlocked