Sounds great on paper but

[Tsardoz]

I tried "a person" with image2 and mask2. This is the best I got:

[omriav]

Hi,

It was never claimed in the paper that this method will be able to generate a plausible output for any inputs.
In this example, adding a person to the desired place is very hard, if you use a different image that a person can
more naturally fit - you will get a better result. For example:

Input image:

Input mask:

Predictions:

In addition, generating people is a hard task in general as you can see.