Could potentially be a compression artefact but I freely admit I’m playing devil’s advocate right now. Do not go down that route we’d end up at function approximation with randomised methods and “well intelligence actually is just compression”.
I actually kind of looked at (jpeg) compression artifacts, and it’s indeed true to the extent that if you compress the image bad enough, it eventually makes it impossible to determine if the color was originally flat or not.
(eg. gif and dithering is a different matter, but it’s very rare these days and you can distinguish it from the “AI noise” by noticing that dithering forms “regular” patterns while “AI noise” is random)
Though from a few tests I did, compression only adds noise to comic style images near “complex geometry”, while removing noise in flat areas. This tracks with my rudimentary understanding of the discrete cosine tranform jpeg uses*, so any comic with a significantly large flat area is detectable as AI based on this method, assuming the compression quality setting is not unreasonably low
*(which should basically be a variant of the fourier transform)
I recreated most of the comic image by hand (using basic line and circle drawing tools, ha) and applied heavy compression. The flat areas remain perfectly flat (as you’d expect as a flat color is easier to compress)
But the AI image reveals a gradient that is invisible to the human eye (incidentally, the original comic does appear heavily jpeg’d, to the point I suspect it could actually be chatgpt adding artificial “fake compression artifacts” by mistake)
there’s also weird “painting” behind the texts which serves no purpose (and why would a human paint almost indistinguishable white on white for no reason?)
the new ai generated comic has less compression, so the noise is much more obvious. There’s still a lot of compression artifacts, but I think those artifacts are there because of the noise, as noise is almost by definition impossible to compress
Could potentially be a compression artefact but I freely admit I’m playing devil’s advocate right now. Do not go down that route we’d end up at function approximation with randomised methods and “well intelligence actually is just compression”.
I actually kind of looked at (jpeg) compression artifacts, and it’s indeed true to the extent that if you compress the image bad enough, it eventually makes it impossible to determine if the color was originally flat or not.
(eg. gif and dithering is a different matter, but it’s very rare these days and you can distinguish it from the “AI noise” by noticing that dithering forms “regular” patterns while “AI noise” is random)
Though from a few tests I did, compression only adds noise to comic style images near “complex geometry”, while removing noise in flat areas. This tracks with my rudimentary understanding of the discrete cosine tranform jpeg uses*, so any comic with a significantly large flat area is detectable as AI based on this method, assuming the compression quality setting is not unreasonably low
*(which should basically be a variant of the fourier transform)
I recreated most of the comic image by hand (using basic line and circle drawing tools, ha) and applied heavy compression. The flat areas remain perfectly flat (as you’d expect as a flat color is easier to compress)
But the AI image reveals a gradient that is invisible to the human eye (incidentally, the original comic does appear heavily jpeg’d, to the point I suspect it could actually be chatgpt adding artificial “fake compression artifacts” by mistake)
there’s also weird “painting” behind the texts which serves no purpose (and why would a human paint almost indistinguishable white on white for no reason?)
the new ai generated comic has less compression, so the noise is much more obvious. There’s still a lot of compression artifacts, but I think those artifacts are there because of the noise, as noise is almost by definition impossible to compress
