Project Pokemon, Not Too Much, Mostly Faces.

Projecting into a GAN is searching for the representation of an image in that model. If you look hard enough you can find close matches for almost any image, even ones that don’t bear much resemblance to the model, but my favorite results are when you search just a little bit. This feels a lot more like asking the model, “What would Bill Hader look like as a cat?”

It’s hard to see because it’s black and black but he has full cat ears here.

Or Asking, “What would the Live Action Film Adaptation of This Looked Like?”

Does it even have to be a face?

Emoji

Grinch Casting Call

Terrible Spiderman Projects (That Feel So Right)

Alignment Is Important

If the images are faces but the eyes and mouths aren’t in the places the model expects:

Assorted Extras

It's all about those early parts in a StyleGAN2 projection. In most you can freeze on early frames and each one feels like a human representation of the target. pic.twitter.com/va7HjcsdAH
— Jonathan Fly 👾 (@jonathanfly) December 31, 2019

Human facial expressions carry so much weight that a StyleGAN2 projection feels like a story is happening off screen. My mind keeps trying to work it out from their reactions. pic.twitter.com/UnG9pQb8gc
— Jonathan Fly 👾 (@jonathanfly) December 30, 2019

The many paths of a StyleGAN2 Projection pic.twitter.com/CHwye92jkf
— Jonathan Fly 👾 (@jonathanfly) January 2, 2020

I'm searching for a latent that matches Kandinsky's "Circles in a Circle" in a StyleGAN2 model trained on nothing. I'm not sure what to expect with this but four hours in we have achieved a sort-of circle! (Official projection code, not the encoder.) pic.twitter.com/J19lMqagj0
— Jonathan Fly 👾 (@jonathanfly) December 22, 2019

Try it yourself:

StyleGAN2 comes with a projector now https://github.com/NVlabs/stylegan2 and you might also want to explore https://github.com/rolux/stylegan2encoder