The algorithm presented here is simple and interesting. Pixel luminance, chrominance, and illumination chrominance are all histogrammed, and then evaluation is simply each pixel's luminance voting on each pixel's true chrominance for each of the "memorized" illuminations. The model can be trained generative by simply counting pixels in the training set, or can be trained end-to-end for a slight performance boost. This algorithm's simplicity and speed are appealing, and additionally it seems like it may be a useful building block for a more sophisticated spatially-varying illumination model.