Generating images is interesting because you can use the difference between the original and the last generation to bias the search, which can cut down the numb

Generating images is interesting because you can use the difference between the original and the last generation to bias the search, which can cut down the number of strokes you need to test. On the other hand, what's happening here is also slightly different to what people usually mean by genetic optimisation in that it looks like it's only the most recent brush stroke that's being randomised. Normally you'd keep the entire stack of strokes around as a genome and allow any of them to mutate, which means you'd need to do a full render at every generation, as opposed to just blitting a single stroke per generation.