If one could write a fit function that returns a value correlated to the meaning of such random image, some genetic algorithm could be used. Also, this would be a perfect project to finally nuke copyright on images. Could also work with random texts, code, etc. Even better, a random piece of code could be evaluated on a simple "does it compile" basis. Hmm, I better get to work.
Along the same lines, read Borges' "The Library of Babel", from 1941.
Quoting from the Wikipedia description: "Despite — indeed, because of — this glut of information, all books are totally useless to the reader, leaving the librarians in a state of suicidal despair."
Obviously this guy and his readers know nothing about math. Th number of possible images (400 by 400 with 24-bit color) is in fact 16,777,216^(400*400), which is an outrageously huge number. You can reduce this slightly by accounting for symmetry.
I wonder if this would be explorable for some much tinier subset of the space of images. Like maybe 32 x 32 greyscale JPEG images with some limit on compressed size that keeps it to interesting images.
I'm specifically suggesting JPEG because it is a compression method based on human perception, so if we focus only on easily JPEG compressed images, we should eliminate a lot of the image that look like snow on a TV, and restrict more to images that look vaguely like something.
Unfortunately, I don't remember enough details of how JPEG compression works to make a reasonable guess how many possible JPEGs this is, and whether we are getting into numbers small enough to be feasible.
A very quick approximation seems to indicate it still isn't feasible. I saved 2 32x32 JPEGs using Gimp at quality 30. One is a cutout of a face from a photo, the other is random noise generated using the filter in Gimp. The resulting JPEG files were 427 bytes for the face, and 525 for completely random.
Now, I have no idea how much of that is the header. So, to help guess that, I saved a solid white 32px image with the same settings, and it was 164 bytes.
