Today's paper is one that fits into a very rare category where I'm prepared to say : this should not have been accepted by the referee and should be withdrawn.
So far as I can tell, what they appear to have done is to wantonly abuse the power of generative AI and not bothered to do proper by-eye data visualisation. Their claim is that they can take input images from ground based telescopes and "enhance" them to the same quality as that from space-based instruments, i.e. improving the resolution with no penalties except making the images a bit grainier. They make this claim repeatedly, saying that this is far faster and cheaper than space missions. Well, yeah, but in terms of the Holy Trinity of faster, better, cheaper, they've completely and utterly sacrificed the middle one.
Worse, they even say :
The generative AI is used as a complex filter that enhances weak signal to turn it into clear image details... While the human eye might not always be sensitive to the high dynamic range enabled by the FITS format, the information can be used by the AI model to identify subtle patterns in the galaxy shapes that can be observed clearly when using space-telescope images.
These are not inherently implausible claims. It's absolutely possible to have robust statistical tests to validate and discover information hiding in data that's hard to see. The problem is that you don't get something for nothing. Just like with stacking, if you want to increase sensitivity, you have to degrade resolution; conversely, if you increase resolution, you lose sensitivity (all other things being equal).
What are they claiming is essentially to have found a clever workaround for this. Unfortunately this is impossible, and their solution really does amount to "making shit up". What they do is to have a training sample of galaxies observed at both low resolution with ground based telescopes, and high resolution with Hubble. These use these as training data for an AI model, which learns the statistical properties of what the low resolution objects look like when properly re-observed with a genuinely more powerful instrument.
The problem is that all the similarities are inherently and purely statistical. There's no physics here : all their model does is create a brand new image of an object that resembles the low-resolution observations on the large scale and has sharper features which are statistically consistent with real objects. There's absolutely no modelling of the evolution of individual galaxies. They haven't "enhanced" the original data at all, they've used it as a rough guide and constructed new images which look similar but with all the details – all the important bits where science happens – being completely fabricated.
Suppose you have a small thumbnail of a galaxy and you want to make it higher resolution. If you don't care about any of the details, then you could absolutely feel free to make a painting which uses this as a guide. You'd get the major details right, but there's no way you could get the exact placement of every star or gas cloud correct. You could use statistics as a guide, yes, e.g. by telling you how many globular clusters you expect, where you expect the star formation to be... but you'd never call your painting an "enhancement" in any sense of the word. Yet this is almost literally what they've done here.
There also doesn't seem to be any point to this. By all means, an AI which can infer detailed properties of a galaxy from data trained in such a way may well have novelty and genuine value : indeed, similar techniques are already used routinely when real data is missing. But you wouldn't say "this galaxy definitely has this particular star formation activity", because you wouldn't know for sure without real observations. You wouldn't bother making an image because that doesn't tell you anything new. It might be nice for public outreach, or because you need a new desktop background, but you could not possibly do any science with it. You would be 100% preventing yourself from discovering any novel features – the whole point of doing research ! – because your images are literally guaranteed to be compatible with objects which are already known.
And they have further fatal failings. In figure 6, they show a comparison of the original low resolution images, their AI-slop, and the genuine Hubble images. But some of the low-res images here are shown with a lousy data scaling, such that all faint features are completely suppressed and invisible. The apparent success of the AI in recovering the galaxy is almost certainly not due to some hidden magic, but simply because the original raw data shows a lot more than they're letting it. Data visualisation matters, people !
Not that the AI images even look that impressive. There are clear differences of detail, sometimes really quite egregious ones, between the real and fake images, exactly as you'd expect when your data analysis technique is "make shit up". What anyone is supposed to do with this is damned hard to understand.
"Here, I've enhanced this image for you so it's nice and sharp now."
"Oh great, thanks ! Wait, so you did new observations, right ? This is real data I can trust and extract new information from ?"
"Umm, no."
Useless. Bloody useless.
Finally, they measure some of the standard major parameters of the AI galaxies and compare them to what they measure from the real high-res images. Great ! And the differences are small, so that's good. But then, I wouldn't expect them to be large because that's the nature of statistics... and far worse, they don't do the same test on the original ground-based images. So this tells us exactly nothing about whether they're somehow recovered new information or not. It's entirely possible the results using the low resolution images would have been just as good as in their so-called "enhanced" images.
Look, there's a nugget of a good idea in here. Actually, to be fair, two good ideas. Training AI on large samples of galaxies to recover features you can't easily see directly is intriguing : absolutely, it might be possible to do this; human vision is amazing but it's not perfect. And their are legitimate artistic uses here. If you have some particular system you want to show to the general public but the original data aren't suitably photogenic, then this method is no worse – and arguably a lot better – than traditional and extremely difficult artist's impressions*. If they'd marketed this as an Automated Artistic Assistant, I wouldn't have any problem with it.
* Before anyone whines about taking jobs away from artists, many smaller institutions simply can't afford to pay artists anything anyway. We don't all have NASA or ESO funding levels.
But that's not what they've done. They're claiming that this will somehow help with new scientific discoveries. This is worse than useless, it's actively misleading and just plain wrong. Now I use LLMs daily with some enthusiasm, so I have no anti-AI axe to grind here... but the claim that this will give us Hubble-quality data for near-zero cost is so ridiculous that this paper should have been released on 1st April. It's awful.
