Top panels show recovery of γ_lens in the presence and absence of distribution shift, when all light is included. We can see that the bias on γ_lens is significantly lower in the absence of distribution shift. The bottom left and right panels show the same when lens light is subtracted from the LSST-quality images. We can see that in this case γ_lens inference is not only less biased but also more accurate in the absence of distribution shift. Error is defined as predicted posterior mean minus truth. The red line indicates perfect recovery.