x.com/horayo328518

moss

Activation by laoshi

either fizgig is an ass trainer or the model has a tendency to consistently increase loss at 1mp. any lr tuners try to dump it all the way to e-6 so the adapter never learns, and even resuming training at 1mp after 0.25mp initial epochs produced worse results. Thus, this draft lora is trained on 512x512.

everyone are doing the new hype model
i wanna be trendy too...

anima

I realized civitai has free storage. Now picking good epoch is your problem.

Activation tag @laoshi wo hao e