I casually posted this post in the Beta Forum of Video A.I…

Since I am currently working on it, I wanted to present it here.

I post this because interesting things come out of Processing when applied to non-stro photos.

In astrophotography, multiple images of the same celestial object are often taken in succession. This is done to reduce noise and other disturbances that may be caused by the camera. When these images are then stacked, the signals from the celestial objects are combined and amplified, while the noise is reduced.

Stacking images can reveal fine details of celestial objects that would otherwise be obscured by noise or other disturbances. Additionally, weak signals that would otherwise be undetectable can be amplified.

Stacking images is typically done with specialized software that automatically aligns and combines the images. The individual images are shifted so that the celestial objects are precisely aligned in all images. Then, the stacked image is optimized using special algorithms to achieve the best possible result.

Stacking images helps to reduce noise, enhance details, and reveal faint features that may not be visible in a single exposure. Stacking software uses algorithms to analyze each image and determine which pixels contain signal and which ones contain noise. The final stacked image will have a higher signal-to-noise ratio than the individual frames, making it easier to see faint structures in the object being imaged.

This means that these images or this output is actually exactly the one you want to have for training.

I wanted to evaluate this today by doing an astro-processing with a normal object (village from some distance).

At night but also during the day.

In relation to the topic, I noticed that, for example, artemis makes the dark areas lighter when denoising and everything disappears somewhat in the average.

Based on my old knowledge, this is not bad, but now I know that it can be done differently.

Single ISO 200 image (30 seconds of Exposure) - Unprocessed but Profiled by Adobe Camera RAW.

Full Astro-Processed image. (124 Stacked images with data (so-called lights), 130 Bias (Sensor Readout noise images), 50 Flat images (Without data, to calculate out the lens darkening, and the matching dark flats (50 images) (to get the readout noise from the flats)

There are also darks, which are made with the same exposure time as the lights, but no light falls on the sensor.

All in all, quite an effort for a single image.

Processed with Siril-Beta and Photoshop (32bit to 16bit conversion + Streching (Editing)), denoised with NoiseXterminator ← Astro AI denoiser.

I must emphasize that I am still in an early stage.

This is the output from stacking (32 bit).

This is a undenoised ISO 200 images, but its processed.


DxO - (XD)

Both DxO and DenoiseRAW are not suitable for any kind of Astro-Processing workflow.

1 Like

I processed this image using astro-processing.

It was de-noised and sharpened with PHAI Enhance Standard.

It’s just an test image so its not beautyful.

Take a look at the shadow area.
This image is not a HDR.


Thanks, TPX. Right down my (next) alley.

1 Like

This is an 12 Minute ISO 100 Exposure.
Edited with Affinity Photo (build in Astro Stack function), Photoshop (Color, Contrast), PHAI (Starpness, Noise)

EOS R3, ISO 100, 12,5 Minutes Exposure) (25 sec per Frame, 30 Frames Stack), EF 300 4.0 @ f7.1.

Un-edited single image in Capture One.

Not sure I understood the point you were making… I take astro images (lights + all calibration frames) and having done the stack, background and star colour calibration and initial histogram stretch with Astro Pixel Processor I put the exported tiff file through Topax Denoise AI. I find the low light option works best. I then import to Capture One for further manipulation (I don’t have Photoshop) with resetting the black point the most powerful tweak. I often send the file for final processing to Topaz Photo AI which selects noise reduction on Autopilot setting and I think the image that is returned is better still. But I do wonnder if Topaz have enough Astro training images to do a specific Astro Image Denoise/Sharpen option? Sometimes the “generic” processing leads to artifacts and stars which look like miniature crystals with jagged edges… I think due to to much sharpening. So Denoise is good, but could it be even better for Astro images?
Stacking increases the signal linearly with number of stacked images but the noise only increases with the square root of the number of stacked images, hence the Signal to Noise ratio improves with stacking.

Project finished.

Half a year of work and planning.

Milky Way with 5 images, each 30 minutes exposed (Sigma 50 mm Art @ f4.0, & Astro Tracker)
Ground with 10 images each 4 min exposed. (EF 200mm 2.0 @ f7.1)

Camera EOS R3 @ ISO 800.

Programs used:
Astro Pixel Processor
Affinity Photo
TopazLabs Photo AI

Pixel Dimensions:
18428 x 8584

1 Like

Astrophotography seems to be an increasingly popular use case for Photo AI. We definitely have this on our radar and are discussing potential features specifically for this type of work.

Thanks everyone for your insights.

The problem, exposure time, which astrophotography has, is quite extensive.

The best way to illustrate the problem is between two images.

My version has about 30 minutes exposure (15 images stacked, 2 min exposure per image at ISO 800), it may be that ISO 800 was ineffective and that is why there is comparatively little to see in my image.

The image below, which I would call good, has 1 hour exposure per image.

The lower image has twice as much signal as mine, the problem is that the signal we want to have does not exist at the 30 minutes exposure.


Another problem with TPAI is that it de-noises the colors of the stars.

But I would like to keep them.

1 Like

Can you upload a sample image where this is especially apparent?

can you send me a link for upload, then you can try for yourself. :slight_smile:


Or you look here.

Denoised with TPAI - Enhance (2x enlarge denoising) and NoiseXterminator.
And Processed a little bit.

Denoised with TPAI Normal @ 3/4/48 - Not Processed