#Stable Diffusion execution time
crisp valeBOT
sturdy sleet
When u use the UI in a111y a lot of things happening in the background
particularly the computer is already started, u already launched the gui, u already loaded model to cache (maybe), etc etc
vs serverless needs to start up a new computer to do all that
potentially if u have lets say back to back requests an already active worker that has spun up and loaded everything would be able to respond faster
also a111y has overhead in starting a server etc in the background
Havent tried this myself but thats why lighter weight libraries that just access the model directly can be better:
Or if u find a cog can use that too
sturdy sleet
Yea hard to tell, are u getting execution time from runpod or from a like python time.time just around the pure part where it is executing? Could be that the time.time part around execution function could be the same.
as i said, there could be many reasons to optimizing a111
b/c at the end of the day it was a webgui to make things easier for ppl > so there is just naturally latency there
I dont think serverless is inherently slower tho i think is the question since ur comparing generating between web ui vs serverless; cause I've seen ashelyk say he got great times due to fastboot (meaning u need a stream of requests to take adv of this) + also ive also seen my more optimized runpod code also respond very fast.
So i think its more of an optimization of a111 issue