>>17331

The process isn't exactly difficult, its just tedious. I did it with ai-toolkit & renting GPUs on runpod cause it takes forever even on a good GPU.

Basically like:

- Find ~10 videos of the thing you want to make (3-6s long). They have to be similar enough that it won't just confuse the training.

- For each video, make an accompanying text file with the same name that describes it and has a trigger word if you want one (so vid_1.mp4 + vid_1.txt)

- Crop each video to focus on the theme of the LORA (easy in clipchamp)

- Optional: if you want the "blink" type loras to use with I2V, add a ~0.3s image of the SAME person/character to the beginning of the video (this is the hard part for a lot of training videos, you need an image of them *not* doing the thing in the video)

- Upload all the videos & text files to a dataset in ai-toolkit

- Configure a training job (there are tons of settings that are tweaked for GPU, lora strength, etc. Wont get into that)

- Train the lora (For a ~2000 step training, this takes about 2.5 hours on the super high end runpod GPUs)


As for doing horses, I *could* try that, but I think there might be too much motion/extraneous stuff for it to work well. I might try like horse cock BJ first since thats a bit simpler motion wise. Plus other existing BJ LORAs can help push it in the right direction