>>76866
Essentially: Lumina has an SDXL sized DiT (AKA newer version Unet), along with the 16 channel FLUX VAE, and finally Gemma2_2b and a tiny CLIP providing text encoding. I fully expect there to be mixed precision downconversions of Gemma2_2b, as otherwise it's half the size of the checkpoint.
My testing of it has shown that it is about 1/2-1/3rd the speed of SDXL, which is better than other DiT architectures I've tried. That said, I don't see it replacing Noob for a lot of people, as I don't expect it to suddenly understand non-booru subjects, concepts, or camera angles just because it's a new model. My car-make/year and (upside down) tests went poorly, with the former not understanding the prompted car at all and the latter having a fair amount of body horror. It did, however, do a good job of putting things like the correct color of prompted shirts onto the correct subjects.
Time will tell. It seems like Newbie is going to have a 8:1 ratio of Danbooru to e6, so we'll need model merging wizards to do their thing again, somehow, and magic together models suitable for /trash/.
Also cute