Hey great read. I was wondering if you could answer something that's been working in my head. If I want to fine tune a very specific small style model, like for instance a specific variety of studio photography, say closeups, very clean softbox linger, in the style of a few different photographers that shine in that field--couldn't it be possible to train on a relatively low number of dataset images with proper captions, say 400 images, uncostely, say 90mn training sessions (and maybe 20 sessions, to get the model really easy and accurate)..? I'm trying Kohya_ss for this, but have been hearing a lot about EveryDream. Issue is I feel I need a gui, being a visual person. But at the same time, I could see myself hiring specialists to more aptly arrive at the custom fine tune. I know my market well, but maybe it is too far out of my way to build good models. Not sure though, as my use cases are very particular. I might use the models as Lora's in generating for my clients. Don't know. Also right on the money, to touch in synthetic datasets as we don't know yet where we are headed legally in the next year. Would love your thoughts on the above thanks

Expand full comment