>>105967897
When I experimented with joycaption a few weeks ago doing a workflow that did illustrious -> flux img2img, I found the phraseology in joycaption to be very repetitive. It tended to make very similar comments in every image. Stuff in your caption like "evokes a sense of wonder" and "otherworldliness" and "commanding presence" and "touch of beauty" I remember seeing over and over. That may partially reflect the kind of fantasy-themed images I mostly make. If a model were trained on its outputs I think you would really need to imitate the joycaption style to get the most out of it.