This computer vision model was trained on available spectroscopic data prior to major DESI releases (includes DESI early data release) -- that is to say, it was trained on (and evaluated with) a biased sample of galaxies. We are combining data from most major spectroscopic campaigns, including auxillary programs in SDSS E/BOSS that do not follow the standard targeting of giant red elliptical galaxies. This has some unknown effects: while it is well known that neural networks do not extrapolate well, it is unknown exactly how well they interpolate in low-density areas of feature space.
Rather than limit what the user can request, we have opted to allow our network to perform inference at any location within the footprint of our telescopes-- even if the pointing does not have a discernable galaxy, if the galaxy would be considered outside our training sample, if the galaxy is blended with another object, or if the galaxy is obscured heavily by galactic dust from the milky-way. WE HAVE PLACED THE RESPONSIBILITY TO YOU, the user, to consider the context of the cutouts when evaluating whether our predictions should be trusted. If you can not discern your target galaxy in the center of the image, then our neural network will not provide an accurate photometric redshift.
That being said, many pointing do not have any discernable flux in the Galex:UV bands. That is normal and reflective of our training set. Its a good rule of thumb to make sure that the galaxy is visible in the optical bands atleast.