Resolution sensitivity test results for the base CNN model and its variations. Left panel: R₂ scores of the best-performing model for each tested resolution ($\sqrt{\text{Total pixels}}\in 3,5,11,25,65,125$), with example “mass” images shown below. Right panel: Same test applied to images smoothed with a Gaussian kernel density estimate, using either Silverman’s rule of thumb (Silverman 1986) or a bandwidth equal to the pixel size. Example “mass” images for the former are shown below. The chosen resolution (dotted purple line and shaded purple region) provides an optimal balance between predictive accuracy and numerical stability.