Machine learning tool scores astronomy images using Claude AI

I’m letting some astro buddies test it out, and if they think its helpful I’ll open it up more broadly.

This tool does not do any processing. It analyses a given loaded image (which is not retained) against publicly available images on certain parameters.

It scores the image on those parameters, and gives feedback as well as a quantitative score.

It’s not being “trained” on any copyrighted data, it doesn’t scrape any data. Its an analysis and feedback tool.

Once I have some feedback from some guinea-pig astro-buddies, I’ll open it to the community here to try out.

Tony Gondola · May 13, 2026 at 02:46 PM

There are things in art that go beyond just the technical aspects that AI can assess. You can make a technically correct image that leaves most people cold and you can make an image that’s technically imperfect but yet it fills the viewer with joy and wonder.

I completely agree that art is subjective — different works affect different people differently, and that's precisely its power. Which is also why most people would hesitate before anointing something as poem of the day or painting of the week. A personal reaction to art is just that — personal.

But IOTD is doing something more than responding to art. It's a collective label applied by a self-selecting group to what constitutes, in their opinion, the best astrophotography. That's no different from what the great academies used to do — and the art world moved on from that presumptuousness long ago. The Impressionists weren't considered art. Neither were the Cubists. You get the drift.

This tool is simply a mirror. It analyses images against a set of craft parameters — signal, processing, composition, detail — and compares them against what a group of people have already selected as IOTD, TPN, and general images. It makes no claim to capture any emotional response to an image (which you’ll admit will vary across every observer); it analyses craft, not feeling. To the extent you believe IOTD reflects expert taste and judgment of what is art, that's already baked into its comparison set.

What's interesting is when that mirror shows inconsistency — when images anointed IOTD score lower than others that weren't. Some will call that bias, others arbitrariness. Either way, the mirror is just reflecting what's already there.

For me personally the score is beside the point. What I value is the systematic, parameter-by-parameter feedback — specific, consistent, dispassionate. I'll then decide what to do with it. That's the beauty of having a tool rather than an arbiter.

YMMV.

Eric Gagné · May 13, 2026 at 08:35 PM

Vin,

I am in no way judging wether this is right or wrong or if it should be done or not or whatever, this is not for me to decide.

I understand what you said in your last post but to me the question remains. What is the point exactly ? Are you using this to compare your work with that of others to decide how your image of target xyz scores compare to others ?

I am honestly puzzled as to what purpose your tool serves exactly.

My motivation is simple: I want feedback on my images that is systematic, consistent, and precise - feedback on craft rather than opinion.

The tool analyses an image against a set of parameters and uses other image sets (IOTD, TPN, general population) as a reference - not to rank my image against theirs, but to generate structured, specific suggestions for improvement. A consistent, structured rubric.

As a byproduct, it also reveals something interesting: the preferences implicit in what gets selected as IOTD or TPN, and the inconsistencies that emerge when those selections are held up to a consistent rubric. That's an incidental curiosity rather than the point of the tool - but perhaps an illuminating one for some.

So in short: better, more consistent feedback on my own work - that's it.

Vin · May 13, 2026, 10:03 PM

Eric Gagné · May 13, 2026 at 08:35 PM

Vin,

I am in no way judging wether this is right or wrong or if it should be done or not or whatever, this is not for me to decide.

I understand what you said in your last post but to me the question remains. What is the point exactly ? Are you using this to compare your work with that of others to decide how your image of target xyz scores compare to others ?

I am honestly puzzled as to what purpose your tool serves exactly.

My motivation is simple: I want feedback on my images that is systematic, consistent, and precise - feedback on craft rather than opinion.

The tool analyses an image against a set of parameters and uses other image sets (IOTD, TPN, general population) as a reference - not to rank my image against theirs, but to generate structured, specific suggestions for improvement. A consistent, structured rubric.

As a byproduct, it also reveals something interesting: the preferences implicit in what gets selected as IOTD or TPN, and the inconsistencies that emerge when those selections are held up to a consistent rubric. That's an incidental curiosity rather than the point of the tool - but perhaps an illuminating one for some.

So in short: better, more consistent feedback on my own work - that's it.

I guess the bast it can do is to see if your image is technically in line with imaging winners.

@Kevin Morefield thanks, this is exactly the kind of structured thinking I was hoping this would prompt - thank you. I'm hoping you'll be willing to share more feedback once you have a chance to play with it?

Your point about conditional scoring is well made and something I want to build in - scoring the TPN→TP transition separately from the general→TPN transition, which would surface the conflicting drivers you describe. I too am curious as to what that might show - it's on my to-do list.

On inconsistency, yes agreed that IOTD judges can only pick from what is put before them and Claude has no visibility into what else was available for promotion on a given day. So this tool isn't an attempt to undermine IOTD, it’s just an attempt to apply a consistent rubric for feedback. One of the things I've observed from learners in other domains is the frustration or confusion that can often come from inconsistent feedback proxies. And that's one of the big motivations for this tool, just to apply a consistent rubric. On any inconsistency it reveals about IOTD or TP selections themselves, that’s just about whether the craft parameters of selected images are themselves consistent (a narrower but perhaps still interesting question for some).

Currently it looks only at the image. I've deliberately kept equipment and metadata out for now - partly because the tool is meant to be a craft mirror, partly because the backyard-vs-remote CDK asymmetry you describe is real (and well discussed here on Astrobin) and I didn't want to bake that in yet. But yes it would be fascinating to iterate the tool to use separate remote vs backyard reference sets.

On temporal drift - completely agree. Claude's training will skew toward more recent images anyway, and I guess that's inevitable as consensus taste and preferences (which is what the reference sets essentially present?) will evolve.

On composition - yes, it's in the rubric. But it's the parameter I trust least for now, with the reason you name being one of the factors. What's interesting is how Claude describes its composition assessment — it tends to fall back on classical framing rules, which may or may not reflect what IOTD judges actually weight.

The art question is one I was coming back to initially but I'm actually more comfortable with it now. This tool doesn't try to label whether something is "art", and perhaps that's the most honest thing about it.

@Tony Gondola yes that's a fair and accurate summary - and for me that's the most useful part of it because it provides such an analytical summary in a consistent, data-driven way

@Gilmour Dickson I hear you, and I think the concern is completely legitimate. To me, a child's crayon drawing has just as much artistic legitimacy as a Rothko (there's a reason Picasso said "It took me four years to paint like Raphael, but a lifetime to paint like a child"). But I'd gently push back on the framing here - this tool doesn't judge art, and it certainly doesn't replace human connection. It's a private feedback instrument, used by a photographer on their own image, for their own purposes. No different in spirit from using a histogram, a noise analyser, or any other diagnostic tool.

Nobody has to use it. Nobody's images are being judged by a machine without their consent. And the humans on AstroBin, including but by no means limited to the IOTD judges and pickers and submitters whose selections form part of some of the reference sets, are very much still in the loop.

The question of whether to ask an AI or a human first is a good one. For some things, like emotional response, humans are irreplaceable. For a structured, repeatable craft rubric at 2am when you're processing an image and nobody's around perhaps a consistent mirror is exactly what many will find helpful? There’s no gospel truth in this tool and that’s not what it claims to be or do (for example some of the feedback it gives on my own images does not work for my own personal aesthetics, but I already know that my aesthetic does not overlap with consensus in some aspects and that’s fine for me).

Sorry this has become a long reply but I wanted to thank each of you for your thoughts. The early feedback from the first guinea-pigs has been positive, so I'm going to share the tool with a few more folks. If the feedback from that is also that it adds value to them in their own processes, then I'll put it out to the broader community. At the end of the day, its aim is just to be a tool - I'm sure the rubrics can be improved with many subtleties and perhaps over time I'll do that. What I don't want to do is over-engineer it because then it implicitly becomes a judgment tool, whereas the point of it is a private feedback tool.

Not working, nothing happens when I select an image.

mmmmm, now I feel dumb but after dropping the file nothing changes, no button to click.

Yeah, I don’t see anything like that. if I drag and drop or left click and select, nothing happens. I’m running Chrome on Win11. I’ll give it a try on a few other browsers to see if it makes a difference.