I am developing an AI-large-model-based denoising software for astrophotography, which is more powerful than Theia AI. The goals are:
Target users and input formats
– Designed for the post-processing workflow of amateur astronomers. Supported input formats: 16-bit FITS (linear), 16-bit TIFF, 8-/16-bit JPEG/PNG (sRGB).
Output and quality standards (acceptance criteria)
– SNR improvement: ≥ 5 dB boost in typical deep-sky (low-signal background) scenes, measured over user-selected background regions.
– Star fidelity: FWHM of bright stars changes < 5 %; no obvious “blob stars”, “blocky stars”, or abnormal halos.
– Faint-structure preservation: nebula edges and dark lanes must not be “wiped out”; qualitatively verifiable in side-by-side comparisons.
– Artifact control: no conspicuous NN artifacts such as repeating textures, grids, rings, banding, or color shifts.
Development has reached the late stage, and I have one question and one request for help:
Question: do people actually need this piece of software?
Help needed: I need unprocessed, original astrophotographs for testing and validation, but such data are extremely hard to obtain. I hope to get help here.
That’s all. I would appreciate your assistance, especially with unprocessed images. My e-mail is
Developing AI denoising software for astrophotography - do people need this?
The capability to handle 32-bit floating-point images exists; it just demands more computing power. Yet the most critical issue right now is that we cannot find any original, unprocessed astrophotography images, so we have no way to verify whether the features we are developing actually work.
Personally I use NoiseXterminator and Topaz DeNoise. Those work fine for me and see no need to purchase an additional noise reduction plug-in. I certainly do not want to take an image out of Pixinsight or Photoshop into a saperate noise reduction software.
My answer would be NO.
Lynn K.
The capability to handle 32-bit floating-point images exists; it just demands more computing power. Yet the most critical issue right now is that we cannot find any original, unprocessed astrophotography images, so we have no way to verify whether the features we are developing actually work.
Download hubble data from MAST, guaranteed predictable and high quality data with plenty of examples of what it “should” look like. The noise patterns may be different due to CCD vs common CMOS now, but this should be something any good denoise software can deal with.
And Photo AI is junk, and also costs money.
If this AI you're making would be free, or otherwise very cheap, then my answer is I want it, not really need it, cuz DeepSNR is already pretty powerful, but i do want it.
edit: you can find raw data in the MOANA errelaz project, but note, stars are six pointed with diffraction spikes, i think, js
Oscar · Oct 18, 2025, 03:54 AM
编辑：您可以在 MOANA Errelaz 项目中找到原始数据，但请注意，恒星是六个尖角的衍射尖峰，我认为，js
Thank you very much for your proposal.
It’s an area that already has some strong AI based contenders, so any new entrant into astrophotography denoising would need to be pretty good, but if it supported 32 bit floating point images and could denoise as well as BlurXterminator, without artefacts, I’d at least give it a look. I think Frank of SetiAstro was giving out a big, unprocessed, high quality, but armature equipment based data set of Andromeda in LRGBHSO this week for people to work with.
Oscar · Oct 18, 2025, 03:54 AM
edit: you can find raw data in the MOANA errelaz project, but note, stars are six pointed with diffraction spikes, i think, js
Excuse me, can the starry sky photos posted on this forum be downloaded? I mean a small batch for development testing, not for commercial use. Is it necessary to purchase them or is there another way?
NoiseXterminator works really well. Is there a reason that we need something else?
John
John Hayes · Oct 18, 2025, 06:24 PM
NoiseXterminator works really well. Is there a reason that we need something else?
John
One reason I can think of is, if you don’t use PI you’re pretty much SOL. If they can match NXT’s performance it would be welcomed by more than a few people.
鬼 狼:Oscar · Oct 18, 2025, 03:54 AM
edit: you can find raw data in the MOANA errelaz project, but note, stars are six pointed with diffraction spikes, i think, js
Excuse me, can the starry sky photos posted on this forum be downloaded? I mean a small batch for development testing, not for commercial use. Is it necessary to purchase them or is there another way?
For that, you need to ask permission from the people who posted the images.
As far as I know GraXpert is also supporting Denoise which works pretty good actually. Before the big NoiseX update I did not see a difference between them. NoiseX was just more aggressive and the results were the same in my eyes.
Unfortunately this changed with the NX update. So a new contender has to be really really really good, because GraXpert as a freeware tool does an excellent job already.
Topaz DeNoise is not discontinued. It is till available for a price. BlurXteminator is a Deconvolution tool. it can reduce noise some, but is not a noise reduction tool. NoiseXterminator is the noise reduction tool, and is also available for Photoshop. The Photoshop version does not have all the adjustment parameters that are available in PixInsight.
Lynn Krizan
Lynn K:
Topaz DeNoise is not discontinued. It is till available for a price. BlurXteminator is a Deconvolution tool. it can reduce noise some, but is not a noise reduction tool. NoiseXterminator is the noise reduction tool, and is also available for Photoshop. The Photoshop version does not have all the adjustment parameters that are available in PixInsight.
Lynn Krizan
https://docs.topazlabs.com/other-apps/legacy
If licenses are still available elsewhere, please name the website.
I have already tested the more recent Photo AI, and it doesn't work the same, its totally unusable for astro.
Lynn K · Oct 16, 2025, 03:30 PM
Personally I use NoiseXterminator and Topaz DeNoise. Those work fine for me and see no need to purchase an additional noise reduction plug-in. I certainly do not want to take an image out of Pixinsight or Photoshop into a saperate noise reduction software.
My answer would be NO.
Lynn K.
I agree with Lynn above.
CS
Jim
I see no reason to discourage the development of a new tool that potentially helps advance the hobby. More choices is a good thing in my opinion. Not every tool works best in all situations, and competition can help drive innovation.
Tony Gondola · Oct 18, 2025 at 08:07 PM
John Hayes · Oct 18, 2025, 06:24 PM
NoiseXterminator works really well. Is there a reason that we need something else?
John
One reason I can think of is, if you don’t use PI you’re pretty much SOL. If they can match NXT’s performance it would be welcomed by more than a few people.
This and BlurXterminator are just two of about 500 really good reasons to learn PixInsight. It is a world class piece of code and it’s hard for me to imagine doing serious imaging without it. Yes…I know that there are other options so save your flames, but in my view, few (if any) have the flexibility, power, or the extensive ecosystem of PI.
John
John Hayes · Oct 19, 2025 at 05:37 PM
This and BlurXterminator are just two of about 500 really good reasons to learn PixInsight. It is a world class piece of code and it’s hard for me to imagine doing serious imaging without it. Yes…I know that there are other options so save your flames, but in my view, few (if any) have the flexibility, power, or the extensive ecosystem of PI.
The topic of whether to pay for and use PI comes up often and the same people respond who are Never going to pay for it for whatever inexplicable reasons…. in the words of the great Strother Martin in Cool Hand Luke, “Some men you just can’t reach”.
andrea tasselli · Oct 19, 2025 at 09:34 PM
Because costs money while the alternative doesn't, that's the reason.
So does a better telescope or camera or mount or… …..
To each his own, it just becomes a tired subject.
John Hayes · Oct 19, 2025, 05:37 PM
Tony Gondola · Oct 18, 2025 at 08:07 PM
John Hayes · Oct 18, 2025, 06:24 PM
NoiseXterminator works really well. Is there a reason that we need something else?
John
One reason I can think of is, if you don’t use PI you’re pretty much SOL. If they can match NXT’s performance it would be welcomed by more than a few people.
This and BlurXterminator are just two of about 500 really good reasons to learn PixInsight. It is a world class piece of code and it’s hard for me to imagine doing serious imaging without it. Yes…I know that there are other options so save your flames, but in my view, few (if any) have the flexibility, power, or the extensive ecosystem of PI.
John
No flames here. I have had a number of PI free trials and with the exception of BlurX and NoiseX, I just never found anything in it compelling enough to make me feel that I needed to have it.
Like with every other invention, in order to succeed, it needs to be better than its predecessors. That can be price-wise or performance-wise. NXT for me, wasn’t even a choice until the recent update. I certainly believe there is room for improvement and I can see how as well. I’d say go for it if you think you can make it.
Oscar · Oct 18, 2025, 03:54 AM
edit: you can find raw data in the MOANA errelaz project, but note, stars are six pointed with diffraction spikes, i think, js
Excuse me, can the starry sky photos posted on this forum be downloaded? I mean a small batch for development testing, not for commercial use. Is it necessary to purchase them or is there another way?
I’m not a great astrophotographer but you can check in my gallery what you need and I will post the raw material in my Google drive. Also let me know if you need the darks, flat and biases with the lights.
AstroStew · Oct 20, 2025 at 09:01 AM
Said like a true PI fanboy 😂
These are the kind of infantile internet comments that ruin a lot of forums. I hope the moderator takes notice before this kind of trolling takes over this one.
I find the development of such a software interesting. I would, however, be less interested in the noise reduction, but in the star shape improvement.
I do wide-angle photography and one common issue is coma etc., especially in the corners. Also, the stars are larger in such lenses.
If one could train such an AI to make the stars of my 16mm lens as crisp as in my 100mm, that would be fantastic! 😏
CS Michael