How to implement the result of Figure 8. in the paper?

[seasonSH]

• background: You can do it by projecting a background image into a coarse feature map first (by changing the inversion code) and replacing the output of background generator with that feature map.
• foreground control: You can control the location and scale of the foreground by controlling the "coords" input into the generator. For example, if you want to zoom in by 1.2x, simply calculate the default coords and divide them by 1.2.

[Chaejongwook1]

First of all, thank you for sharing the good thesis and code. I'm an undergraduate research student. Can you explain the foreground control in more detail? I'm interested in controlling the local latent control Thank you.

[MingCongSu]

Thanks for your great work. @seasonSH However, I got problems when projecting a background image into a coarse feature map. Do you mean like projecting custom background image into the latent space of SemanticStyleGAN? Thank you.👍