You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
They are pretty similar, which means the initial random image is the same.
However, there is some source of unseeded randomness, I think it will be great to locate and fix it :)
I suspect it's due to CUDA's nondeterministic behavior, as it doesn't happen in the CPU mode. Slight numerical imprecision is sometimes inevitable, especially when dealing with half-precision models. A workaround is to "warm up" the model as you did, and after a third call or so it'll more likely emit deterministic outputs.
Activity
dribnet commentedon Oct 3, 2021
Any chance you get better results if you leave off the noise_prompt_seeds? Just curious if that might be the source of it.
altsoph commentedon Oct 3, 2021
I tried it with the same results
robinsloan commentedon Oct 4, 2021
I suspect that CLIP is the culprit!
See: openai/CLIP#13
Merge remote-tracking branch 'sgallag/master' into future
Make output_dir if it doesn't exist
Make all directories for output_dir
renamed output_dir to outdir and fixed default
cog fixes for new outdir setting
outdir templates and save settings.yaml each run