The McCann99 Retinex is a particular algorithmic implementation of Retinex theory. Retinex algorithms attempt to compress the dynamic range of an input image to make an output which can be displayed and has a pleasing rendition. The McCann99 Retinex algorithm employs a multi-grid structure, processing the image in a pyramidal manner from coarse to fine resolutions. This hierarchical approach is designed to enhance computational efficiency and accelerate convergence. In recent work, the structure of the McCann99 Retinex has - by appealing to classical work in numerical analysis - been modified to be a per pyramid level Jacobi iteration and as such is not only faster, but aligns Retinex with other theories of early visual processing (including centre-surround operators and path-based theories). In this paper, we conduct a complementary systematic investigation into how images are down- and up-sampled as part of the multi-grid computation. The McCann99 and Jacobi Retinex algorithms use a simple box-filter approach, which means that down-sampled image data must be aliased. We investigate a set of other commonly used sampling methods, including bilinear, bicubic, Lanczos2 and 3, sinc with rectangular, Hamming and Kaiser windows. We are interested in using image quality metrics to determine whether substituting a different sampling approach into the Jacobi Retinex leads to better quality outputs being generated. Our investigation leads us to an interesting conclusion: while more complex down- and up-sampling algorithms can improve image quality, the simple box method performs well. If image quality is the primary concern, sinc is found to be the best method for down- and up-sampling. However, it requires an appropriate window function to taper the truncation and avoid ringing artefacts.