This page contains some advice on how to get best results using StarNet
Use color images whenever possible
The first tip is very short: when it comes to star removal using StarNet there is no advantage in processing channels one by one. Color images contain more information and allow for more precise star removal.
Stride value
Stride value should be based on the size of stars in your image. For wide-field images, which generally have small stars, I recommend setting as high as possible – 384 in the latest versions.
For the rest of the images the default 256 will do.
There is 99.99% chance you don’t need to go lower than 256, don’t waste your time, unless you actually have a reason.
Try drawing spikes on your stars
This one is a bit weird. Some people report that StarNet misses stars in their images. This seems to be happening with Greyscale images (see tip #1) that have perfectly round fluffy stars. This is due to the fact that my training data has no perfectly round stars, even my images from refractors have somewhat diamond-like shape to them. This will be changed in the future.
I found that one way to fight this problem is to draw small spikes on stars using software like Star Spikes Pro 4. This helps StarNet to identify stars as stars and remove them. However, I should also note that a few missed stars can be removed in PS in a few clicks using Clone Stamp tool or similar.
Use 2x up-sampling only when needed
The 2x option in PI plugin mainly serves one purpose – to improve star removal for images with very tight stars. Here is an example (you might need to zoom in to see what I’m talking about):
The image above is taken using high-end equipment an therefore is both very smooth (has high SNR) and has very tight bright stars. It seems like this scenario is problematic as StarNet generates visible artifacts, as shown on the right: you can spot radiating patterns where the stars used to be.
Another problem found in the same image is StarNet missing tiny stars in complicated areas (bottom center):
The example above actually shows both problems at the same time: radiating patterns around stars in dark areas and missed stars in brighter areas.
Both problems disappear when 2x up-sampling option is used:
This way we can get better star removal, although at the cost of 4x computing time. The only drawback of this method is that big stars get even bigger, so if you have huge stars in your image the performance on them will suffer.
Playing with stretch helps sometimes
Auto STF will put you in a good spot for star removal most of the time. However, that’s not always the case. Let’s take a look at example below:
The image shows a detail of the Veil Nebula, namely The Witch’s Broom and Pickering’s Triangle Nebulae, in Hydrogen alpha. Although the starless image looks good (at the first glance at least), the star mask clearly shows that something went wrong around the Pickering’s Triangle. Let’s zoom in into problematic region:
We can see that StarNet ‘ate’ some of the brightest filaments. One potential way to combat this is to try running star removal on color image instead of a Greyscale (see tip #1):
But what if that’s not an option?
This is another case when 2x up-sampling can be useful: increasing the scale allows the neural network to avoid removing bright high-frequency components of the image. Here is the result with 2x up-sampling:
Much better! However, still not perfect, some of the brightest filaments still got removed.
The next idea is to try to reduce the brightness of the image to make even brightest filaments less bright. Here is the result after moving STF sliders around:
Zooming in into the same region as before (now using auto STF to show the result better):
Just with these two simple tricks we were able to achieve perfect star removal:
One thing to note is that the example above uses linear image. Repeating the same procedure using pre-stretched image might be less optimal, although definitely worth experimenting with if no linear image available.