Face Normalize Data Augmentation

 

I'm working on a facial normalize data augmentation option for gallery show.  It auto-detects the main face in the source image, and then normalizes the position to a standard location.  I'm just doing the center points right now, but i'll add a scale option as well.

So why would you want to do this? Have you ever noticed how those StyleGAN faces are all in a standard normalized central position. Did you ever wonder why they did that?  Think it though carefully.

A few more simple examples of stacking the source faces below in successive gallery show cycles. Using our old friend the paint synthesizer Init QuickEdit.

You can see the effect of not doing the size scaling in that last one.  Which could be fun for artistic purposes. But of course you want that full normalization to a standard face position for other applications.

You have been able to do this kind of thing in a much more tedious way in Studio Artist forever.  Using the Inverse Warp Context action step.  But you had to manually draw the bezier curves for the inverse warp. Just doing it automatically this new way is way simpler.  Although it's going to fail for any source image that the frontal face detector craps out on.

I will also build another automatic one using the facial features.  It will have a very different visual aesthetic to it, since it will be based on a sparse warp.  I need to provide an option to use these in the high level Studio Artist interface as well. So we will add that to the ridiculously long 'to-do' list as well.