Working directly with pixels affords the unique benefit of being able to use images. It also gives a context where the visual output is identical with the data that generated it, i.e. it has one of the most primitive Renders, one of the three pillars.
Starting with an image, we traverse its pixels from left to right, top to bottom. We start by choosing a random length L, and then copy the color of the first pixel L times. At step L+1, we choose a new length and copy the current pixel's color. If all lengths are 1 we recover the initial image, play with various lengths to see the results for yourself.
Here we take 2 or 4 source images and interleave them so that each little 2x2 pixel block features pixels from each. If we used an all-white and all-black image, this would generate a very fine checkerboard. We can change the block size to make a larger checkerboard and make something of a quilt out of chunks of the images.
If you haven't looked at this process already, head on over to the diffusion page.