>>470172271 more
>>47018326you'll only notice a difference on a card old enough that it can't use fp16 at all or is significantly slower at fp16 so webui bumps it up to fp32
GTX1080 tier. when you have a gigachungus fp32 + EMA weight model...it still runs it at fp16 precision anyways. you'd have to manually launch the bigger model with "--no-half --precision full" to see the difference, if there even is one.
>excess dataEMA weights are used during training for...stuff
