@gelato_al_pollo one example of "2 bytes per pixel" https://stackoverflow.com/questions/9972869/dxt4-texture-compression

the color space of the RGB channels was sacrificed to drop the image size by 1/3rd, and one of the channels ends up with more precision than the others.

similar stuff is done in audio. mp3 and opus are based on "psychoaccoustic" models where they basically make a computer model of how a human hears most sounds and then they re-prioritize the bits around the important bands.