It occurs to me that you could do error diffusion during the HAM-conversion stage. This would probably help break up the 3-pixel wide horizontal bands that occu

It occurs to me that you could do error diffusion during the HAM-conversion stage. This would probably help break up the 3-pixel wide horizontal bands that occur during a big color change, causing a different component to switch first in successive scanlines. Specifically, imagine that you diffuse all error _down_. Changing from (0,2,3) to (31, 31, 31) you'd change R first, then G, then B. But with a portion of the error (0, -29, -28) added to the white pixel of the 2nd row, you'd change G first, B second, and R third.