I've been using your Vodka model for Stable Diffusion for the last two years, on and off. I like it. I have trained a lot of Stable Diffusion models myself - small ones, Hypernetworks and Embeddings.
To your point of teaching people about AI, I think having people, gather, clean, and label 20-50 images, and then having them create their own hypernetworks, LORAs, and embeddings is a great teaching method. It's really fun to make your own model, and then try and create things with it in Stable Diffusion. Then, you can go back and forth and see, "why is my model only spitting out these types of images, why can't it do this, how is it doing this?" It's fun and pretty educational to understand how these things work.
I've used claude a good bit, too. It's really amazing technology. But, when you use it a good bit, and if I compare it to the small SD models i've made, it's pretty easy to see how narrow the LLMs, and the smaller models, can be sometimes.
I could not agree more, and it's so interesting that only now with your comment do I see it. What you described is pretty much our experience as well and you gave me lots to think about. Training a lora on a small set of image-caption pairs is possibly the simplest, most instructive and most accessible ML task out there. Thanks for the comment!
I've been using your Vodka model for Stable Diffusion for the last two years, on and off. I like it. I have trained a lot of Stable Diffusion models myself - small ones, Hypernetworks and Embeddings.
To your point of teaching people about AI, I think having people, gather, clean, and label 20-50 images, and then having them create their own hypernetworks, LORAs, and embeddings is a great teaching method. It's really fun to make your own model, and then try and create things with it in Stable Diffusion. Then, you can go back and forth and see, "why is my model only spitting out these types of images, why can't it do this, how is it doing this?" It's fun and pretty educational to understand how these things work.
I've used claude a good bit, too. It's really amazing technology. But, when you use it a good bit, and if I compare it to the small SD models i've made, it's pretty easy to see how narrow the LLMs, and the smaller models, can be sometimes.
I could not agree more, and it's so interesting that only now with your comment do I see it. What you described is pretty much our experience as well and you gave me lots to think about. Training a lora on a small set of image-caption pairs is possibly the simplest, most instructive and most accessible ML task out there. Thanks for the comment!