Conditional biases of the shear point estimator ĝ_i as function of true properties of the variable PSF, in equipopulous bins. The same training and validation datasets are used for all panels, but three different ML setups are shown. Top panels: ML is not informed about about the particular PSF of each galaxy. Middle panels: field coordinates (x, y) are used as input features. Bottom panels: shape parameters ε₁, ε₂, and σ of the PSF are used as input features. As in Fig. 9, only cases with ⟨S/N⟩≥10 are considered for this unweighted analysis of the point estimates.