AI & Creative Tools
Quantisation, Explained: Getting a 20GB Image Model Onto a 12GB Card
Every current image model ships in several sizes, and the labels — fp8_e4m3fn, Q4_K_M, int8 ConvRot — look like they need a maths degree. They do not. Here is what each format does to the numbers, why outlier channels are the thing that breaks naive quantisation, and how to pick a build for the VRAM you have.