Claude AI on integration time limits for astrophotography

Interesting, I never thought about calibration, specifically flat fielding accuracy as a limiting factor with ultra long integration times. Over all, the numbers it’s coming up with seem to be in the ballpark.

The0s · May 10, 2026 at 02:51 AM

Interesting… I'd be curious to know how it came up with the 0.1-0.5% for the error from flat frames (I doubt those numbers are easily found on the Internet) and how the point of diminishing returns would change for a more light polluted location. In any case, this is a cool topic, so thanks for sharing!

Google fabian neyer night sky flats.

i am in the middle of building a set of flat areas for my luminance. I shortly talked about it in the description of one of my latest pictures.

Scott Badger · May 10, 2026 at 11:02 AM

On the other hand, images like the recent M104 IOTD shows what can be done with integration time far beyond what most consider ‘diminishing returns’.

Just came here to say the same thing. Returns might diminish but there are still returns to be had for the dedicated

John Hayes · May 10, 2026, 06:37 PM

Jerry,

Your question touches on a number of interesting things that I’ve given a bit of thought to in the past. First, I’ve experimented a bit with AI for solving some slightly mathematically challenging problems. A year ago I gave Claude and ChatGPT the same thermodynamics problem that I had worked out myself in some detail. At that time, Claude really struggled and after directing it through multiple revisions (around 6x) I finally gave up. It just couldn’t get there. On the other hand, ChatGPT did a much better job. I only had to give it 2-3 course corrections before it provided what appeared to be the correct solution. I recognize that within the last year, Claude has been much improved; however, the lesson is deeper than that. And here’s the lesson: If you don’t already have a pretty deep understanding of what the solution should look like—and why, you can get completely fooled by an answer provided by AI. To be clear, that not to say that AI isn’t a useful tool. It is but you need to be very cautious to probe the answer and you need to use other means to confirm that it makes sense.

In this case, Claude has pointed out something interesting. Years ago I first posted a calculation that expanded on the “Rule of Five” published in “The Handbook of Astronomical Image Processing 2nd Edition”, by Berry and Burnell (now out of print). My calculation (posted in other threads here on AB and on CN) showed how dark calibration (using stacked dark masters) affects noise in a stacked image, but that result might not be quite the same as how flats affect stack statistics. As Janesick points out in “Photon Transfer”, signal variation due to FPN varies directly with signal strength and can far exceed photon noise for some sensors—and that’s one of the three key reasons that we use flat calibration. But the issue that Claude has mentioned goes beyond simple image calibration. It extends to how the SNR is affected by FPN when you combine calibrated and dithered frames in the stack—and it’s not clear to me that the answer that you got from Claude is correct. For one thing, Claude doesn’t show its work and the details of why it is saying what it is saying are vague. Dithering is a key component to reducing the effects of FPN so it is a very important factor when you consider any limitation on SNR with respect to total exposure time.

I’ve got too many other things I’m working on at the moment to dig more deeply into this but perhaps when I get some time, I can give it a bit more thought. In the meantime, I would recommend approaching those results with some caution.

John

Thanks John! I am definitely skeptical of AI’s response, especially when I know the imagers I respect and admire are getting excellent results with very long integration times. Even using a refractor similar to my 130mm aperture, I see fine images that are much longer than AI says is the “point of diminishing returns”.

Charles Hagen · May 11, 2026, 06:01 AM

There are a couple issues here, though the main issue is the premise of the question. To be extremely pedantic, “diminishing returns” start at sub #1. Every subsequent sub-exposure contributes less and less to the stack than the one before. The curve begins to flatten before it even starts. What you are really wanting to know is “at what point is it not worth it for me to gather more data?”, the answer to which is entirely subjective.

In a vacuum, stack SNR will increase by 41% (sqrt 2) for every doubling of integration time regardless of equipment or light pollution. (yes there are some very subtle exceptions to this and it is complicated by read noise and changing conditions, but for our purposes it’s close enough) This means that if you double your exposure time, the proportional change in SNR will be the same regardless of how much time you’ve already contributed. Obviously, however, doubling your total integration quickly becomes untenable. While we’d all love to have 512, 1024 or even 2048 hours of integration per target, most of us aren’t so patient.

If you are sitting at 100 hours of integration and you are unsatisfied, it is best to ask yourself if you’d be happy doubling your efforts for a perceptible, but not striking improvement in the noise profile - If yes, carry on... If no, time to devote your efforts elsewhere. While unfortunately it is hard to predict at the beginning of a project, this is the only reliable way to test whether more time is “worth it” or not in my experience.

Hi Charles,

Makes sense. I haven’t yet gone beyond 60 hours or so, will probably extend that to around 100 hours at some point. I took the AI answer with a grain of salt because I’ve seen so many outstanding images with integration time in the hundreds of hours.

Arun H · May 11, 2026, 04:56 PM

I have found it to be an excellent learning tool. I am sharing this to just give you guys a sense of how powerful it has become.

LLMs are absolutely powerful tools and you’re that they have progressed a long way in recent times, but they still suffer from training data availability. In your example, you are asking questions that have been well discussed in scientific literature and broadly across the internet, therefore it is more likely to be correct. Niche subjects in massive fields are still fairly common and the data used to train Gemini is (relatively) rich in that kind of data. Our area of interest is more niche and less common across the internet, the training data will contain less relevant context to answer that question with nuance. My point is that extrapolating its apparent expertise in one field onto another field will often give you a false idea of its real capabilities.

I've found that there are certain types of questions that AI answers very accurately. Others not so much. But the real strength of AI is its ability to learn from itself.

I asked Claude a question about classical music harmonic theory and it got it wrong. As soon as I explained the conditions that must exist for a particular type of harmonic structure to be identified as such, it immediately corrected itself. If only people were that quick at self-correction!

Arun H · May 11, 2026, 04:56 PM

I suspect AI has come a very long way in one year. I have had some pretty detailed math and physics problems I have posed to AI and it seemed to provide answers that were verifiably correct.

I’ll comment on the actual subject of this thread later in order to stay on topic, but as for this point:

I recently graduated from a good University with a degree in physics. AI from 2 years ago (so most definitely AI today) could do nearly all the problems a physics undergraduate is presented with. This is because physics undergraduate work, and even a significant number of the problems that you’d find at graduate level, are well documented and studied because that is the purpose of science. We have established such a strong base for students to learn from, that an AI has no issue scraping the internet and providing a solution (with sources) for these problems. Occasionally it makes an error but it is almost always easily corrected by another prompting. Does this mean AI can do physics? No, as the actual research involves asking new questions, something I’ve yet to see AI do effectively. I could devolve this topic into a post about how undergrads are just poisoning themselves by not learning the material or whatnot but that’s been covered extensively elsewhere. The point here is, yes AI can do these “difficult” physics problems because we’ve already solved them and understand them.

As for the actual subject matter of the thread:

IMO the AI gives a result that is annoyingly close enough that I don’t want to bother correcting it, but still wrong. I would parrot basically everything Charles Hagan, John Hayes, and Scott Badger said. Most notably I find that the AI has no distinction between variety of targets within each of the categories presented. Building off my rant above and some personal assumptions, I don’t believe there the depth of information of astrophotography integration time as there is on physics, so it’s no surprise the AI answer is lacking.

Edit because I always think of something else to mention shortly after I hit post:

My post is a bit redundant as the above few posts cover the ideas I’ve mentioned, though I would like to clarify that the danger in using AI for questions like Jerry’s here is that the AI is built to return an answer and does not have the means to confirm the answer if there is not a clear solution already present.

Arun H · May 12, 2026, 11:16 AM

Most humans would struggle at comprehending basic principles such as the above!

As for the original question of integration time and Claude’s response. I think many humans would struggle too, in the absence of good source data.

I believe we are mismatched in what standard we’re holding the AI to. I am not suggesting the average person is studying college physics, nor a physics research. Likewise, I am not suggesting the average person is an experienced astrophotographer. I am suggesting that it is too often that the flaws in an AI’s answer are forgiven because of (in this case) lack of adequate reference material and because it got more basic and verifiable information correct. I have seen far too many examples, both on the internet and in my personal life, of people “developing theories” of dark matter, the big bang, or whatever their interested in because they trusted an AI. It is almost always the case that these people verified lower level information just like you have done with the E&M problem. Frankly, I do not care if it was an AI or a person presenting the original attached PDF or their new theory of everything. The confidence in which it is incorrect and the difficulty for someone ill informed to detect where the information strays from truth is a massive issue. Unfortunately AI has made cases like this much more common, and in turn much more exhausting for all sides involved to debunk.

To be clear I do not intend any hate towards Jerry or someone using AI to ask questions, but I do encourage a more formal learning avenue, especially when you’ve past well established topics.

Arun H · May 12, 2026, 11:52 AM

The problem was taken from an undergrad physics textbook, and so too the solution, both of which were posed in an exam at a reputable university. So perhaps the problem is that lower level information is being peddled in our undergrad textbooks and universities?

This is straying too far from on topic imo so it will be my last comment in this thought line.

If you reference the holy grail of E&M textbooks, Introduction to Electrodynamics by David J. Griffiths, it contains information such as “A+B=B+A” (page 2 chapter 1, in this case it is establishing vector addition), while also containing information on Electrodynamics and Relativity, which one might consider a bit more advanced.

It is essential to cover the basics, perhaps one might go so far to call them the essentials, as they are the groundwork from which we have developed modern physics. The problem arises when you take someone from that second year undergraduate class, who’s just done well enough to get an A, and present them as a reference for a new theory. Do not misrepresent understanding of the basics as ability to answer more complex questions.

Anecdotally, I would say the question you posted is not incredibly “low level”, though definitely undergraduate level. When my class saw questions like this most everyone was halfway through their second year of college and already taken math classes that, not too long ago, were not offered widely. I believe we were just under halfway through the E&M class when we would see a problem similar to this.

Edit: Is not incredibly “low level”. Oops.

C.Sand · May 12, 2026, 12:16 PM

If you reference the holy grail of E&M textbooks, Introduction to Electrodynamics by David J. Griffiths, it contains information such as “A+B=B+A” (page 2 chapter 1, in this case it is establishing vector addition), while also containing information on Electrodynamics and Relativity, which one might consider a bit more advanced.

I must be getting old. In my day, “Classical Electrodynamics” by J.D. Jackson was the holy grail of E&M. I have to admit that it’s been so long since I took that class (as a physics grad student), that when I look back at my copy, I can’t even understand any of my margin notes much less any of the material. Use it or lose it! :))))

John

John Hayes · May 12, 2026 at 11:43 PM

In my day, “Classical Electrodynamics” by J.D. Jackson was the holy grail of E&M.

I am not a physics major, but I did take Classical Mechanics (Goldstein book) in grad school and loved the material. I bought Griffiths a few weeks ago to go deeper into E&M and its relationship to relativity - believe it or not, I asked Gemini what book it would recommend and it recommended Griffiths and Jackson, warning me against the denseness of Jackson. I will admit that I have no desire to go after it. That book seems to have a reputation. This Amazon review of Jackson had me laughing:

📷 image.png