Chromapair

The problem

Matching an image’s color to a reference is usually done by eye, one image at a time. It rests on judgment, so the result cannot be measured, repeated, or reused.

The approach

Chromapair learns a fixed transform from the areas that line up between a source image and a reference. It tries two kinds of model, three 256-entry lookup tables or a regularized RGB polynomial, and keeps whichever has the lower error under four-fold cross-validation, measured in CIEDE2000.

It does not depend on what the images show or how they are laid out. A smaller source can sit inside a larger reference, several separate areas can be fitted at once, and the tile size used for averaging is chosen for you.

The result

What comes out is a validated JSON file that can be applied from the command line or from Python, to one image or a whole batch. A color decision becomes something you can version, share, and use again.