P
ponder
@ai.ponder.ooo
alt acct of @ponder.ooo where i plan to post more ai stuff, mostly stable diffusion but sometimes other models.
alt text will be of highly inconsistent quality
11 followers15 following195 posts
unet has to be at fp16 precision to fit on my gpu nicely. doing the rest of the calculations at fp32 yields a barely-discernible difference, but doing them at fp64 causes a dramatic shift in final result
64-bit vae instead of 32-bit vae for the decode does not significantly change the resulting image but makes it waaaaay slower
P
ponder
@ai.ponder.ooo
alt acct of @ponder.ooo where i plan to post more ai stuff, mostly stable diffusion but sometimes other models.
alt text will be of highly inconsistent quality
11 followers15 following195 posts