>>105967384
I made a tool to help with building loras from video stills:
https://huggingface.co/quarterturn/facesaver

I've used it with my molmo captioner (preferred because it is the least censored) to pretty easily crank out lora training datasets.