Astronomers have released AI-enhanced images of 63,202 galaxies, using paired Hubble and ground-telescope observations to teach a model how blurred structures can look at higher resolution. The catalogue could help researchers choose targets from wide sky surveys for closer investigation.[1]
In Monthly Notices of the Royal Astronomical Society, Sai Teja Erukude and Lior Shamir describe a conditional generative adversarial network trained on paired galaxy images from the Hubble Space Telescope and the DESI Legacy Imaging Survey. The team released a catalogue of 63,202 enhanced galaxies, as well as the software and paired training material. The goal is to expose structural features in broad ground surveys that are difficult to discern in the raw images.[1]
A training set in which both cameras saw the same thing
The paper assembled approximately 20,000 pairs. For each target, one image came from Hubble observations of the COSMOS field, and the other from a ground-based survey centred on the same sky coordinates. This paired arrangement is the critical teaching signal. The computer is not simply being told what a pretty galaxy looks like; it is comparing a lower-detail observation with a higher-detail recording of the same object.[1]
The team tested formats derived from JPEG and from FITS, the data format widely used by astronomers. FITS can retain high-dynamic-range intensity information not obvious in a display-ready image, which may help identify subtle structures. Images were processed into consistent dimensions for training. The model then learned a transformation from the ground view to a representation resembling the associated space view.[1]
A conditional generative adversarial network contains a generator that proposes an enhanced image and a discriminator trained to distinguish generated output from a true target. A reconstruction term also rewards agreement with the paired Hubble image. The model can consequently learn regularities in how spiral arms, bright cores and extended galactic structures appear when the observation is clearer.[1]
The scientific danger of a convincing picture
The model draws on both the light recorded in an input and the regularities it learned from other galaxies. Where the observation is ambiguous, a sharp-looking arm or faint companion may be an inferred feature. That is why an enhanced image is useful for forming a hypothesis, while a discovery claim still needs observational support.[1]
A practical use is triage: flag a galaxy with an unusual apparent shape, return to the original survey data, then seek higher-quality observations if the feature warrants them. The enhancement helps decide where to spend scarce telescope time; the original pixels remain the measurement record.[1]
What could 63,202 sharper candidates change?
The catalogue is valuable as an experimental resource. Astronomers could use a consistent enhanced view to flag unusual galaxies for deeper observation, test whether morphology classifications improve, or evaluate which structural claims survive comparison with real space imagery. Code and paired data make such checks more feasible than a product demonstration showing a few handpicked pictures.[1]
A wide survey supplies breadth; space observations supply detailed training examples. The catalogue’s research value will depend on how well that combination helps astronomers select and classify galaxies against independent observations.[1]
Sources & further reading
- Amplifying the imaging power of digital sky surveys with space telescope data and generative AI Monthly Notices of the Royal Astronomical Society · 22 Aug 2026 · Accessed 24 Oct 2026
- Enhancing Ground Based Astronomy using GenAI Erukude and Shamir / GitHub · Accessed 24 Oct 2026