A statement I often hear on AB that puzzles me is concerning noise. Sometimes the comment is an image has too much noise and sometimes too little. It seems there is a magical middle ground where noise is just right. I think we would all agree that creating detail that doesn’t exist in the object through too much AI sharpening is something to be avoided. On the other hand, making an image where noise is imperceptible is generally a no no. The thing is, noise doesn’t exist in the object, it’s an artifact of our cameras. Why do we go to great lengths to eliminate other non-object artifacts through calibration and even masking yet noise is something that should not be eliminated. Our tools are certainly good enough to get very close to the truth of an image yet this feeling about noise persists. Love to know your feelings on this.
A case against or for noise.
Tony Gondola · Jul 20, 2026, 05:12 PM
The thing is, noise doesn’t exist in the object, it’s an artifact of our cameras.
The dominant form of noise is not a camera artifact but arises from the statistical nature of photon incidence on the sensor, which follows a Poisson distribution. It is unavoidable and cannot be improved by use of better cameras.
I look at it similarly to why we like to see movies at 24fps, an artifact from the film days, rather than 60 or greater. Certainly the later is more like reality, but it’s not what we’re used to and it’s referred to as the Soap Opera Effect. Even Peter Jackson filmed the Hobbit at 48fps, but the audience thought it looked cheap and jarring.
We expect some underlying noise in space shots since, as Arun points out, some is unavoidable.
Personally, I magnify to 200% and will noise reduce so that I still see noise there. When viewed at 100%, there’s enough noise to my brain thinks it looks natural, but not enough that it detracts.
Arun H · Jul 20, 2026, 06:15 PM
Tony Gondola · Jul 20, 2026, 05:12 PM
The thing is, noise doesn’t exist in the object, it’s an artifact of our cameras.
The dominant form of noise is not a camera artifact but arises from the statistical nature of photon incidence on the sensor, which follows a Poisson distribution. It is unavoidable and cannot be improved by use of better cameras.
That’s interesting. So even if we had noise free electronics and perfect amplifiers we would still see some degree of noise in our images?
Rick Krejci · Jul 20, 2026, 07:08 PM
I look at it similarly to why we like to see movies at 24fps, an artifact from the film days, rather than 60 or greater. Certainly the later is more like reality, but it’s not what we’re used to and it’s referred to as the Soap Opera Effect. Even Peter Jackson filmed the Hobbit at 48fps, but the audience thought it looked cheap and jarring.
We expect some underlying noise in space shots since, as Arun points out, some is unavoidable.
Personally, I magnify to 200% and will noise reduce so that I still see noise there. When viewed at 100%, there’s enough noise to my brain thinks it looks natural, but not enough that it detracts.
Interesting way to look at it. That’s a pretty good guide I think.
Tony Gondola · Jul 20, 2026, 07:11 PM
That’s interesting. So even if we had noise free electronics and perfect amplifiers we would still see some degree of noise in our images?
Yes. The best video that explains the various sources of noise is by Robin Glover. I don’t have a link handy, but you should be able to do a web search. A much more technical writeup on the various sources of noise is this 2008 paper by Emil Martinec that I read over one Christmas a long time ago which was well worth my time. It is still relevant.
https://homes.psd.uchicago.edu/~ejmartin/pix/20d/tests/noise/
In the context of astro work you aim, when possible, to make shot noise the dominant term by increasing sub exposure length so it drowns out other sources of noise in each sub - the logic being this is the source of noise that you cannot control and is unavoidable, but you can make so that the others are much smaller in comparison.
To your original question - take a look at some of the best images here. There is very fine grained noise in the dark regions. This makes them look natural and pleasing. Compare them to overly denoised images where the background looks like a glossy sheet of plastic. Some noise is very much desirable.
Arun H · Jul 20, 2026, 07:35 PM
Tony Gondola · Jul 20, 2026, 07:11 PM
That’s interesting. So even if we had noise free electronics and perfect amplifiers we would still see some degree of noise in our images?
Yes. The best video that explains the various sources of noise is by Robin Glover. I don’t have a link handy, but you should be able to do a web search. A much more technical writeup on the various sources of noise is this 2008 paper by Emil Martinec that I read over one Christmas a long time ago which was well worth my time. It is still relevant.
https://homes.psd.uchicago.edu/~ejmartin/pix/20d/tests/noise/
In the context of astro work you aim, when possible, to make shot noise the dominant term by increasing sub exposure length so it drowns out other sources of noise in each sub - the logic being this is the source of noise that you cannot control and is unavoidable, but you can make so that the others are much smaller in comparison.
To your original question - take a look at some of the best images here. There is very fine grained noise in the dark regions. This makes them look natural and pleasing. Compare them to overly denoised images where the background looks like a glossy sheet of plastic. Some noise is very much desirable.
Thanks Arun, I’ll take a look at both sources.
Astrophotography can be fairly conservative in it’s outlook in terms of adopting new things. It might seem strange to hear that because we expect it to be high tech considering what we are trying to do.
If you want an example, a lot of the advice you can find out their for exposure times and what not is still based off of old CCD’s with horrid read noises and pixel bleeding. I think you can add noise to this as well, because there was just no avoiding it with both the noise injected into the final stack by CCD’s, and the limited processing tools people had at the time. Now we have the tools and the cameras to allow us to remove most if not all noise without losing detail. The prerequisite to this is having the integration time to support such a move.
That being said, personally I prefer to see either a finely grained noise profile, or none at all if they were able to preserve the details. You can tell when de-noising has been excessive because details were destroyed and things look overly soft. In my opinion, excessive de-noise is not removing the graininess, it is removing details that were otherwise there.
Being a dinosaur from the CCD era, I am both accustomed to seeing a bit of noise even in a world class image… Even hubble images have an amount of noise.
I have certainly seen and know a few people who’ve been doing this since BEFORE CCD’s.. These people will not use AI tools at all on their data, they won’t use any methods of correcting trailed stars, instead, they will agressively cull their data such that only the pristine frames make it into the final image… Are they wrong? no… Do I follow the same rules for my own imaging? also no…
Everyone has a different palette or tolerance for noise in an image.. I feel like when there is none at all in the background or really faint nebulosity/dust, that I’m being lied to… Because no matter how amazing your camera is, no matter how many hours you sink into a target… If you aren’t imaging in light polution, the sky background will have received far less photons than the target, and thus, there will still be noise… So when there is not, I tend to feel like the processing was too heavy handed… Likewise, when there is a boatload of high frequency colour noise in an image, I find that really detracting from the subject… It is, unfortunately, a fine line, and a line I’m guilty of crossing myself…
My image of M78 from 2024 is a good example of an image where I feel I went too hard on the noise supression… The weather sucked for that season, and as much as I wanted to get 24~36h on the target, I managed a total of 9h I think… Not enough to show what I wanted to show, not enough to be able to push the data for the colour and contrast I wanted to achieve. Despite this, I pushed the data pretty hard anyway, and I got what I wanted (mostly), but I also created a very noisy image.. So a few iterations of NXT over the whole image were run, as well as some selectively masked blurs being applied… Overall, I don’t love the result, but it’s better than just giving up on the data…
Noise is subjective, but you’re right, for most, there is a minimum and maximum ‘acceptable’ threshold level, and that will be different for every person.
Having owned all the following cameras:
SBIG ST8XE, ST9XE, ST10XME, STL-11000M, ST-8300M, QHY8, ZWO 1600MM PlayerOne Artemis C (IMX294), ToupTek ATR2600M (IMX571) and ToupTek SkyEye45AC Plus (IMX366).
I’ve seen it all when it comes to standard noise profiles… From incredibly clean through to horrific… I’ve seen massive blooming on stars through to the advent of anti-blooming gates on CCD wells, through to completely overflow free CMOS sensors.. Even amp glow so serious that without darks you can’t make out your target through to sensors that for all intents and purposes do not need darks for significant reason at all…
I have a pretty high tolerance for noise and other camera based artefacts that may exist in images… But when I see one of this silky smooth images with zero noise, I feel that they always look like they are wrapped in cling wrap, and the stars look out of place in them…
The main thing I’m getting from these comments is that the whole question is very subjective with a lot of factors coming into what each person’s ideal noise profile is. Chemical photography was like that too. Some people just loved grain and would develop Tri-X in Rodinal just to get more. Others would swear by Panatomic-X where the grain was almost at the molecular level. At the end of the day it’s aesthetic choice like many other choices we make in the workflow.
One thing that’s not been mentioned is that just a touch of noise can enhance perceived sharpness when the image is viewed full screen with the effect falling apart once you start looking at the image 1:1 There’s a very interesting interplay going on here that I don’t fully understand but find interesting.
Noise is what uncertainty looks like. Photons crossing trillions and trillions of kilometers, hitting a millimeter-sized sensor that someone put there by sheer spacetime luck. I expect to see noise.
I also have heard it said that one should leave some noise in the image because it “looks sharper”.
To be brutal, I call total BS on that.
…… But not because it is untrue, it does give an illusion of added sharpness, but the important part of that is the word “illusion”. I don’t think we should be creating illusions.
IMHO, the proper way to do noise reduction is to zoom in to an area of the image that has (or should have) the finest detail available in that image/object and use preview on/off to look before/after the noise reduction. Then maximize the NR while minimizing any loss of detail.
It is pretty clear that what you should see is the same real detail with as w/o the noise reduction. The actual amount used is not relevant, just whether it decreases detail or not.
If you discriminate against an image because it did not leave as much noise as you “expect” , you could just be shooting it down because it had great resolution or great seeing or lots of integration time. These are not things we should be doing!
Arun H · Jul 20, 2026, 06:15 PM
The dominant form of noise is not a camera artifact but arises from the statistical nature of photon incidence on the sensor, which follows a Poisson distribution. It is unavoidable and cannot be improved by use of better cameras.
True, but it is still not part of the object, which is what we are trying to capture. The effects of that are always gonna be there but can be reduced by exposure time, aperture, skies, etc.
What really bugs me is when people make an assumption about how much noise “should” be in a given image w/o taking the above things into account, which will vary a great deal from one image to the next.
Rick Krejci · Jul 20, 2026, 07:08 PM
I look at it similarly to why we like to see movies at 24fps, an artifact from the film days, rather than 60 or greater. Certainly the later is more like reality, but it’s not what we’re used to and it’s referred to as the Soap Opera Effect. Even Peter Jackson filmed the Hobbit at 48fps, but the audience thought it looked cheap and jarring.
We expect some underlying noise in space shots since, as Arun points out, some is unavoidable.
Personally, I magnify to 200% and will noise reduce so that I still see noise there. When viewed at 100%, there’s enough noise to my brain thinks it looks natural, but not enough that it detracts.
The reason films are shot at 24fps is because that’s what best matches our eyes with respect to motion blur. Correct me if I am wrong, but this is what I’ve learned making drone videos.
The quality and percentage of noise reduction matters.
Noise reduction heavy enough to appear smeared or creating the appearance of phantom shapes seems to be the main complaint against heavy noise reduction.
People often remove actual data as well. Some of the “noise” in wider shots is stars and people just blur them away to clouds.
I think there are a few things going on with the idea of too much denoising.
Completely denoising is essentially introducing artifacts because the denoising will not get us to the answer of what the object would look like if there were no noise present; but rather what the denoising process guesses is there. And I don’t mean only AI guesses.
Completely denoising produces smoothed details that destroy some of the resolved features
But the bottom line is that it looks unnatural to the eye. And that’s true in terrestrial landscape photos too.
The “right” amount of denoising produces a balanced result that preserves as much of the resolved details as possible without leaving the noise level distracting. A low but present level of noise is not noticeable unless you go looking for it. Maybe more importantly it needs to be consistent across the image. For example, if you have a really low signal OIII shell and obliterate it to a soft blob but maintain the sharpness present in the Ha it’s not going to look right to the eye.
Noise/Graininess is at the pixel level. Yet images are frequently processed, reviewed and judged at screen sizes and NOT examined at the intrinsic platescale or more to assess whether an image is overly smoothed out. There may be exceptions- but I am stating a general experience I have had. (A related issue is that many images are processed to look good at screen resolutions and not at the intrinsic platescale (or near the seeing limit) of the data.)
A good test: Show images with and without noise reduction applied at (or near) the intrinsic image size. If you cannot tell the difference between the images or you have to squint to see it- it probably isn’t overly done. It will be obvious if it is overdone.
I think this topic is better when constrained by what exactly is being looked at. Since this is a pixel-to-pixel variation- you really need to be at these scales to judge and images would need to be made available at the scale necessary to tell.
I agree with Kevin that global processing is desirable. In my own processing I try to “earn” the right to display the faintest signals. I might see more in my data because my eye-brain is very good at seeing stuff in the noise- but if it is simply too noisy to display I choose not to instead of attack it with some algorithm that makes contiguous pixel values. Basically I assign a threshold in my images. For those that like Columbo…. “This far… and no farther.” The way noise is handled I guess really is the measure of a man.
Arun H · Jul 20, 2026, 06:15 PM
Tony Gondola · Jul 20, 2026, 05:12 PM
The thing is, noise doesn’t exist in the object, it’s an artifact of our cameras.
The dominant form of noise is not a camera artifact but arises from the statistical nature of photon incidence on the sensor, which follows a Poisson distribution. It is unavoidable and cannot be improved by use of better cameras.
I realized there is a bit more to this than simply saying the use of better cameras will not result in an improvement.
What we are after is SNR; and SNR improves as the square root of the total gathered signal.
Therefore, improvements in QE should allow us to get to improved SNR in a shorter period of time. Hence the use of better cameras will not help with total photon shot noise (and in fact will increase the total noise in a given period of time, since it is proportional to \(\sqrt S\) where \(S\) is the signal) but will certainly help with improved SNR.
Adam Block · Jul 21, 2026, 06:47 AM
Noise/Graininess is at the pixel level. Yet images are frequently processed, reviewed and judged at screen sizes and NOT examined at the intrinsic platescale or more to assess whether an image is overly smoothed out. There may be exceptions- but I am stating a general experience I have had. (A related issue is that many images are processed to look good at screen resolutions and not at the intrinsic platescale (or near the seeing limit) of the data.)
A good test: Show images with and without noise reduction applied at (or near) the intrinsic image size. If you cannot tell the difference between the images or you have to squint to see it- it probably isn’t overly done. It will be obvious if it is overdone.
I think this topic is better when constrained by what exactly is being looked at. Since this is a pixel-to-pixel variation- you really need to be at these scales to judge and images would need to be made available at the scale necessary to tell.
I agree with Kevin that global processing is desirable. In my own processing I try to “earn” the right to display the faintest signals. I might see more in my data because my eye-brain is very good at seeing stuff in the noise- but if it is simply too noisy to display I choose not to instead of attack it with some algorithm that makes contiguous pixel values. Basically I assign a threshold in my images. For those that like Columbo…. “This far… and no farther.” The way noise is handled I guess really is the measure of a man.
Presentation size indeed something I struggle with. In regular photography I really don’t worry much about what might be seen by someone viewing the print from an inch away, it’s the overall impression at normal viewing distances that I care about, the overall presentation of an image as a complete statement.. There’s so much pixel peeping that goes on with astrophotography that makes me feel that I do have to be concerned although I’d rather not be
I totally disagree with those that say denoising should always leave some easily noticeable noise at a 1:1 viewing scale.
I have been doing imaging since the days of the ST4 in 1993 and have seen many other arguments that use similar logic as a basis for various conclusions as to what is “off limits” in processing. These arguments always come from the same source and that is the optical and physical science involved in the imaging process. To the scientifically trained person (and many imagers have that background), that seems to be an unassailable position.
But what seems to one person to be an unassailable position can seem to another to be an inappropriate conclusion. I am in the latter category and think we need to step back and take a broader view of just what we are trying to do when we produce an amateur esthetic image (and that is what 90% + of the images on Astrobin are). It is not science data and has not been from the time we completed basic calibration (and sometimes before that). My goal is to present the maximum detail in the object but also to make an image that is pleasing to the eye and informative to the viewer. Noise is neither pleasing nor informative.
This is most apparent for me in images of small planetaries taken at the limits of the available resolution. I do a lot of those with a plate scale of .3 arcsec under seeing that is often in the single arcsec range. In those situations the fine single pixel scale “detail” is often not detail at all but rather mostly noise. The true structure shows up only at scales that span at least several pixels. This is what you try to bring out with sharpening and contrast enhancement. The single pixel level variations are mostly noise or at least cannot be clearly identified as real structure. If they cannot be identified as real structure during processing, they must be assumed to be artifact (I define noise as an artifact in this sense because despite its real physical nature, noise is an artifact in the sense that it is not a real part of the object). The goal is to be left with real object detail as best you can determine by comparing the stretched but otherwise unenhanced image to the image that has been enhanced and denoised.
I almost always find that you can preserve 95% of real object detail and still remove 95% of the single pixel scale noise and you have a much more esthetic and informative image as a result. NOTE: when I say 95%, I am not referring to the denoise settings, but the visual result.
Will an image with some pixel scale noise left look sharper at magnifications less that 1:1? Yes, it will, but that is a visual illusion and I would point out that this could also be achieved by over sharpening and no one is suggesting that as a good idea.
At the end of the day everyone needs to do things they way they think is appropriate and that is fine. The only real problem may be when the images are being evaluated for awards which inevitably introduces whatever ideas the evaluator may have regarding processing. This is why I think that two things should be done during evaluation: 1) Examine at 1:1 and 2) Compare to other good images of the same object, also at 1:1. I know that is more work, but is also more accurate.
Arun H · Jul 21, 2026, 02:47 PM
Arun H · Jul 20, 2026, 06:15 PM
Tony Gondola · Jul 20, 2026, 05:12 PM
The thing is, noise doesn’t exist in the object, it’s an artifact of our cameras.
The dominant form of noise is not a camera artifact but arises from the statistical nature of photon incidence on the sensor, which follows a Poisson distribution. It is unavoidable and cannot be improved by use of better cameras.
I realized there is a bit more to this than simply saying the use of better cameras will not result in an improvement.
What we are after is SNR; and SNR improves as the square root of the total gathered signal.
Therefore, improvements in QE should allow us to get to improved SNR in a shorter period of time. Hence the use of better cameras will not help with total photon shot noise (and in fact will increase the total noise in a given period of time, since it is proportional to \(\sqrt S\) where \(S\) is the signal) but will certainly help with improved SNR.
…and read noise is so low as to almost be below consideration. For my setup it runs about 0.6 Ev.
Bill McLaughlin · Jul 21, 2026, 03:10 PM
The only real problem may be when the images are being evaluated for awards which inevitably introduces whatever ideas the evaluator may have regarding processing. This is why I think that two things should be done during evaluation: 1) Examine at 1:1 and 2) Compare to other good images of the same object, also at 1:1. I know that is more work, but is also more accurate.
I would absolutely agree with that and it’s really what prompted me to post on the subject.
Adam Block · Jul 21, 2026, 06:47 AM
A good test: Show images with and without noise reduction applied at (or near) the intrinsic image size. If you cannot tell the difference between the images or you have to squint to see it- it probably isn’t overly done. It will be obvious if it is overdone.
Not sure what you mean by “intrinsic image size” in this context. If you mean “full screen” on the veiwer’s display then I agree. If you mean 1:1 scale then I totally disagree since at 1:1 all the nasty small scale noise becomes much more apparent.
Of course that brings up a very important point and that is what image scales are available to the viewer and that matters a great deal! As many have said, leaving some noise when the viewer cannot access 1:1 scale might be esthetically fine or appear apparently sharper. OTOH on platforms like AB where 1:1 is often available, what is optimal might be different.
Perhaps some way to control what the viewer can do in terms of viewing scale might be a good feature? (EDIT: There already is).
Kevin Morefield · Jul 21, 2026, 06:03 AM
A low but present level of noise is no noticeable unless you go looking for it.
But I do go looking for it (at least when processing)! 🙂