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
Hello, I think it's one of the best repos with inpainting!)
But I have problems when I'm trying to do equal result.
Predicted img has artifacts will colored lines. I don't know why.
How much images was used for training?
I've used default settings, have trained TF version with 150k on pretrain and 150k other img from Places2 with random rect mask, but the result is bad comparing to paper, why?
"The step of excluding s is omitted for computational efficiency." Is it a critical moment for training equal result?
The text was updated successfully, but these errors were encountered:
Hello, I think it's one of the best repos with inpainting!)
But I have problems when I'm trying to do equal result.


Predicted img has artifacts will colored lines. I don't know why.
How much images was used for training?
I've used default settings, have trained TF version with 150k on pretrain and 150k other img from Places2 with random rect mask, but the result is bad comparing to paper, why?
"The step of excluding s is omitted for computational efficiency." Is it a critical moment for training equal result?
The text was updated successfully, but these errors were encountered: