Very cool idea! The current outputs have a relatively simple effect though, not so far from a simple overlay where the blocks are "active". And the blocks which are not active get restored to the base image again. Maybe one could do selective in painting instead. For example, inpaint the cells which turn non-active in a generation with one kind of prompt, and the ones that turn active with another prompt. So it would pull cells towards two different concepts, but remember a bit the history/state of a cell.