>>107112746
Yeah work whould be done on increasing the resolution and quality of output of existing architectures(well mostly XL) like better vae etc. Quality would increase alot by that alone because now the model would be able to tell that these pink splotches of color in the initial noise are supposed to be two separate fingers and not some gigeresque body horror, it would also make crutches like controlnets more effective due to that reason etc. Yeah 16GB would be the target, it's not quite there but it'll be the new norm soon, basically if a model can't be loaded fully into a 16GB GPU without any humiliation rituals it won't take off for the next few years.