2d psf solution from January 2021 are poor in the red (R1)

I noticed that my solutions from January 2021 are systematically poor in the red part of the red detector. 

poor_fit_detector.png shows the quantity that measures the quality of the fit. It is log10(chi^2 residual/ chi^2 max), where chi^2 max is np.mean(sci_image^2/var_image), and chi^2 residual is np.mean((sci_image-model_image)^2/var_image).

example_poor_fit.png shows the example of the single psf. We can see how the fit does not capture the shape of psf in the data.