Influence of training the magnitude on a differential piston estimation accuracy compared to training on a magnitude range. Differential piston RMSE versus number of photons per frame for the SP25 turbulence configuration, with 13-frame averaging. Results are shown for three regimes within a monochromatic uncorrelated setting: the baseline model (blue, corresponding to Fig. 10), a model trained at magnitude m ∈ [15; 18] (red), and a model trained at m ∈ [8; 15] (yellow). The dashed purple line indicates the noise-free monochromatic limit.