Comment by computator
3 years ago
I thought so too, so I tried to find out if they were scraped from the web. I saved about a dozen generated images and used Google Reverse Image Search and TinEye to see if I could find a match on the Internet. But no luck.
Maybe the cat images look so real as compared to generated human faces because our brain is tuned to recognize human faces and can perceive the tiniest flaws, but not so with cat faces.
I don't think my search proves anything however. With hundreds of millions of cat pictures on the web, I wonder if simple transformations on a real photo -- like flipping or slightly rotating the image, squishing it vertically or horizontally by 10%, adjusting the color -- would be enough to hide the original from Google or TinEye?
>Maybe the cat images look so real as compared to generated human faces because our brain is tuned to recognize human faces and can perceive the tiniest flaws, but not so with cat faces.
Lucky for all of you, I spend way too much time staring at cats and while most of the ones generated here are pretty good, there's quite a few where the proportions are off and they look fake or edited to me. If I saw them outside of the context of being generated images, I'm very sure I'd still be staring at them trying to figure out what was wrong.
I think you should be able to find a latent vector that returns a cat that is part of the original training data (or at least very close to it). Most of the outputs will not be real cats at all though. However, it's pretty simple to try and find the latent vector that reproduces a given image, e.g. https://github.com/NVlabs/stylegan2-ada-pytorch/blob/main/pr...
TinEye can certainly find different resolution (and aspect ratio?) versions of images, and somewhat cropped images. I'm not sure how much rotation it can handle, though.