(obligatory non-milk version for later DeviantArt upload included)Turns out training an AI to label the data to train another AI is turning into a real hassle. It's going well, but still a totally different world. Straying a single inch off the beaten path of Automatic1111's walled garden you quickly realize how firmly he was holding your hand: the world of AI is so much more convoluted, complicated and deep than just diffusion, let alone the made-to-order diffusion UI we've been given on a silver platterHopefully, though: the effort will be worth it. Like all things in computer science: the diffusion finetune, dreambooths and the like follow the rule of 'garbage in, garbage out', which also means enriching your data to be more accurately captioned should have a meaningful effect on the resulting models' performance and flexibility