I considered this idea a couple of years ago - using different EFM patterns to generate a visible picture, perhaps introducing correctible errors to enhance the picture.
It really does work, to some extent. If you make a 30 minute wav file with AC silence at DC-offset 0x1515 (EFM encoding of 0x15 is 00000010000000) and a 30 minute wav file with AC silence at DC offset 0xA8A8 (EFM encoding of 0xA8 is 01001001001001), and burn them on a CD, you will see two distinct areas - The second track has a lot more land/pit transitions, and it does show.
First problem: synchronizing in the along-track direction. Back-of-the-envelope calculation - A typical land/pit passes under the laser in about 1.6 microseconds. At half-CD (R=6cm,) with a reading velocity of 1.2 m/s, one revolution takes 2*pi*0.06/1.2 = 0.37 m / 1.2m/s = 0.3 seconds.
This basically means that there are about 200,000 lands and pits in one revolution, and the writer would have to be precise enough to align the pattern on subsequent passes of the spiral winding. I think it can't do that (but I could be mistaken).
Second problem: I remember finding a Philips patent that covers this. So by all means try this for fun, but don't expect to get rich