Toying with Video generation right now I'm breaking down workflows so that I can have a tight lock on generation. Below are initial failures at training Loras that know what I look like. I'd say 70% there. My dataset was 15 images, the new one is over 900. I thought I'd share what training an Avatar looks like from the backend. You do a lot of training, and play with different models and see where things land. The tinkering part is fun, though it's a lot of waiting. Basically you feed it lots of reference images in different settings and iterate a couple thousand steps of training. Then you can use the LoRA on a model to force your character style/appearance. This is a 2 step process. as I'll be getting the base LoRA setup then applying modifications to turn it into a cartoon. The idea is to experiment with a educational short form video instructor.