Finding optimal data for inpainting is a key problem for image compression with partial differential equations (PDEs). Not only the location of important pixels but also their values should optimise the compression quality. The position of such important data is usually encoded in a binary mask. The corresponding pixel values are real valued and yield prohibitively high storage costs in the context of data compression. Therefore, quantisation strategies for the pixel value domain are mandatory to obtain high compression ratios. While existing methods to quantise the data for PDE-based compression show good quality, unfortunately, they are too slow for many applications. We discuss several strategies, based on data clustering models from machine learning, to speed up the quantisation step.