>>105735804
>99% what this was is just he got a good image at 8 steps and bragged about how good the model is, nothing else.
no, he actually change the model so that it works at fewet steps since v29.5, here's his new process

>>105691124
>hope he can do this in a separate branch while continuing normal CFG training for a while to see how things pan out.
I doubt he'll do something like that, there' already 2 branchs, it means it'll end up with 4 branchs, I suspect he'll do like the distilled low step stuff and merge everything into one model