If you're aiming to minimize MSE, then that's not too expensive and it's similar to gradient descent. Adjust weights/counts in histogram from quality feedback measured from remapped image.

This shifts axis-aligned subdivisions to local optimum (I've tried applying that after every subdivision, but it didn't matter - once in postprocessing is enough!) Voronoi iteration after generating palette (like a final step in LBG quantization).The biggest wins I've found (and implemented in libimagequant) were: Click the Options button on the bottom right of the toolbar and go to Save Frames at the top. However, Wu's method needs lookup tables for the whole color (hyper)cube, which requires either large (or for RGBA still ridiculous) amount of RAM or reduced precision, and losing bits of input is way worse than having suboptimal division. Load the animated GIF into 7GIF by dropping the file onto the window or manually using the Open button/menu. Wu's method is nice in this case because it exactly measures variance of sets after a split, while median cut just estimates. I've tested median cut that can cut at various angles and variants that cut out spheres, and it didn't make much difference.Ĭhoice of the box to cut and the location where you cut is more important. Aligned subdivisions are not that much of a problem actually.

Ffmpeg gif from images