Re:Diffusion Models Beat GANs on Image Synthesis
by yarg on Saturday May 22, 2021 @06:41PM
(#27250900)
Sorry, I missed this - unfortunately I did not save the picture.
Repeatedly clicking refresh on https://thispersondoesnotexist.com/ I eventually reached an image where the top of a generated man's head melted into the room behind it.
Melted, not faded.
The room and the man's head were not in anyway separate and superimposed, it was a perfect fluid transition.
Looking at it, it genuinely felt like my brain was malfunctioning.
I don't know the extent to which it was due to under-training or some more fundamental limitation of GANs; the feeling that I had was that the generator had found an exploitable hole in the discernor's comprehension.
Repeatedly clicking refresh on https://thispersondoesnotexist.com/ I eventually reached an image where the top of a generated man's head melted into the room behind it.
Melted, not faded.
The room and the man's head were not in anyway separate and superimposed, it was a perfect fluid transition.
Looking at it, it genuinely felt like my brain was malfunctioning.
I don't know the extent to which it was due to under-training or some more fundamental limitation of GANs; the feeling that I had was that the generator had found an exploitable hole in the discernor's comprehension.