Russell Croman Astrophotography  

 

 

StarXTerminator

Home | Free Trial | Purchase | Support



Photograph courtesy Adam Block. Hover over the photograph to see the original.


StarXTerminator is a tool to remove stars from astronomical photographs that works in Photoshop and PixInsight. This allows separate processing of stars and background objects, or simply leaving out the stars altogether.

StarXTerminator uses an advanced convolutional neural network with a unique architecture suited to this task. This network has been extensively trained on photographs from a wide range of instruments, from camera lenses to the James Webb space telescope. Small stars, big stars, huge stars, and even diffraction spikes are recognized and removed, with minimal impact to non-stellar features.

The end result is a very smooth elimination of stars, with minimal residual artifacts.

StarXTerminator has a "universal" licensing system: a permanent license for any version (Windows/Mac/Linux, Photoshop/PixInsight) will work for any other version. Permanent licenses can be used on two computers – up to three upon request – provided you are the primary user of StarXTerminator on all of them. This limitation is per-computer, so using StarXTerminator in Photoshop and PixInsight on the same computer, for example, only consumes one license activation.

Usage Notes

Keep in mind that star removal is an extremely challenging problem, and no star removal tool will ever perform 100% perfectly on all images. StarXTerminator was trained to work on images produced with a very wide range of instruments, from camera lenses to the James Webb Space telescope. It will produce excellent results on a majority of images, but occasionally there may be cases where stars are not completely removed, or some minor non-stellar structure is.

StarXTerminator was also trained to handle a limited range of optical aberrations such as minor focus errors, guiding errors, coma, field curvature, etc. It may not function well on images taken with instruments with serious optical deficiencies. Invest the time needed to get your optical setup functioning well, not only for better star removal, but also for the best detail and contrast in nebulas and galaxies achievable with your equipment.

If despite these efforts you find a case where StarXTerminator does not seem to perform well, review the usage notes below. If still getting poor results, feel free to contact support. Perhaps your instrument setup is unique and your data can be included in the training of the next version of the neural network. We continually train the neural network on new data, and RC Astro is always happy to receive suggestions for improvement.

Here are some usage notes and tips for the PixInsight and Photoshop versions of StarXTerminator:

PixInsight

  • Use StarXTerminator as early in the processing flow as possible, ideally right after integration, with the data still in a linear state (i.e., prior to any stretching). This will generally produce the best results, and gives the added flexibility of being able to stretch the starless and stars images separately depending on the desired end result.
  • StarXTerminator is trained on images stretched using a simple midtones transfer function (MTF). When processing linear images, StarXTerminator internally performs such a stretch automatically, then precisely reverses it after processing to return the image to a linear state.
  • The MTF stretch is the same method used by PixInsight's HistogramTransformation tool. Any processing that significantly alters star profiles relative to this method may reduce the effectiveness and/or quality of star removal. In particular, an arcsinh stretch and a generalized hyperbolic stretch (GHS) can create star profiles that are indistinguishable from small elliptical galaxies, and will result in StarXTerminator not removing, or only partially removing, the stars. Other processing operations such as masked stretch, high dynamic range processing, etc., may also alter star profiles enough that they will not be recognized as stars by the neural network.
  • If generating a star image, don't use STF Auto Stretch on the resulting stars image. StarXTerminator will translate the STF parameters of the original image to the stars image to make subsequent stretching easier. Auto-stretch will destroy this STF information and give a false impression of the significance of very faint background residual pixel values.
  • If generating a star image from a linear image, don't select the Unscreen option, which is for nonlinear (stretched) images. Simple subtraction is the best star extraction method to use with linear images, and will result in the best star color accuracy. For best results when recombining the (perhaps separately processed) starless and stars images, do it after stretching both, and use screen blending in PixelMath.

Photoshop

  • Use StarXTerminator prior to any major processing of the image. The effectiveness and quality of the result from StarXTerminator will be adversely affected by any processing that significantly alters star profiles.
  • To generate a separate starless and stars layer, follow this procedure:
    • Start with the image you want to separate on a layer
    • Duplicate this layer twice so you have three copies of the original image
    • Remove the stars from the top-most layer using StarXTerminator
    • Duplicate this starless layer so you have two copies of it
    • Drag one of the starless layers down so it is between the two remaining original layers. Name this layer "Starless"
    • Invert the top-most starless layer and set its blend mode to Divide
    • Invert the next layer down (a copy of the original)
    • Merge the top two layers into one. This will now be an inverted image of just the stars.
    • Invert that merged layer. This is now the stars layer. Name it "Stars"
    • Set the blend mode of this Stars layer to Screen
    • The combination of the Stars and Starless layer below should look exactly like the original image
    • You should end up with three layers in this order (top to bottom):
      • Stars (screen blending mode)
      • Starless
      • Original (in case you need it)
    • Process the Starless and Stars layers separately to taste

System Requirements

Please see current known issues and planned improvements on the support page linked above.

StarXTerminator requires a computer with a modern CPU having instructions required to run neural networks. Older CPUs lacking FMA, AVX, and SSE instructions are not supported. Request a trial before purchasing to make sure it will work on your machine.

  • MacOS
    • Version 11 (Big Sur) or later
    • Photoshop CC or PixInsight 1.8.8-9 or later
    • GPU/Neural Engine recommended for fast performance, but not required. Not all GPUs are supported.
    • Apple silicon (e.g., M1, M1X) supported natively
  • Windows
    • Windows version 10 or later
    • Modern CPU capable of running neural networks (AVX, SSE instruction set extensions)
    • PixInsight 1.8.8-9 or later, or
    • Photoshop CS4 or later (64-bit only)
  • Linux
    • Ubuntu 18.04 or later, or equivalent (glibc 2.27 or later required)
    • PixInsight 1.8.8-9 or later

PixInsight Module Version History

Version Date Comments
2.0.3 30 Sep 2022 Adjusted Unscreen Stars function to avoid saturated pixels when one or more channels in the original image are clipped. Code-signed Windows and MacOS modules for better compatibility with anti-malware software.
2.0.2 15 Sep 2022 All-new AI11 neural network with greatly improved performance on a wider range of images, trained on images including data from the James Webb and Hubble space telescopes. Automatic handling of linear and nonlinear images.
1.3.1 12 Jul 2022 Added batch processing function to save starless and stars images generated from a collection of input images.
1.2.0 8 Jan 2022 Fixes an issue that prevented the "stars" and "linear" parameters from working correctly in scripts and ImageContainers. The stars-only image now inherits the FITS parameters and astrometric solution of the original image. MacOS: fixed a memory leak.
1.1.0 8 Nov 2021 Necessary for compatibility with AI version 6. Uses PixInsight's update repository system for installation. AI files are now installed as part of this process.
1.0.0 8 Oct 2021 Initial release

Photoshop Plug-in Version History

Version Date Comments
2.0.4 1 Oct 2022 Fixed large-overlap mode in Windows version. Added processing progress indicator for Affinity Photo. Adjusted MacOS and Windows installers for compatibility with Mac App Store and Microsoft Store versions of Affinity Photo.
2.0.2 15 Sep 2022 All-new AI11 neural network with greatly improved performance on a wider range of images, trained on images including data from the James Webb and Hubble space telescopes.
1.2.2 30 Sep 2021 Update for Windows version only: eliminates the need to move the three Microsoft DLL files to Photoshop's or Affinity Photo's application directory. These files can now be placed alongside of the plugin file itself (StarXTerminator.8bf).
1.2.1 27 Sep 2021 Update for Windows version only: reverts to running only on the CPU. GPU selection was found to be unreliable. This will be addressed in a future update.
1.2 26 Sep 2021 Supports processing of 8- and 16-bit RGB and grayscale photos. Windows version selects GPU if available. Removed batch size adjustment and calibration (unnecessary). Fixed compatibility with Affinity Photo for Mac.
1.1 18 Sep 2021 Adds native greyscale image processing. Adds AI download for Mac version, and download progress indicators for Windows and Mac versions.
1.0 12 Sep 2021 Initial release

AI Version History

Version Date Comments
11 15 Sep 2022 All-new neural network architecutre trained from scratch on a much wider data set, including data from the James Webb and Hubble space telescopes. Greatly improved retention of non-stellar detail, especially from the glare of very bright stars. Revamped noise matching engine that fills in noise where the stars once were that closely matches the rest of the image.
10 18 Apr 2022 Corrects a malfunction in the noise matching module that could cause black or white squares around the edges of uncropped images or in clipped regions of images. Same core neural network and training as AI8.
9 9 Apr 2022 Corrects a malfunction in the noise matching module that could cause black or white squares around the edges of uncropped images. Same core neural network and training as AI8.
8 18 Mar 2022 Re-trained from scratch for greater precision in removing stars and retaining non-stellar details. Training data set further expanded to include tough cases such as bright stars, crowded star fields, and especially galaxies and galaxy clusters. Particular attention given to retaining galaxy disk details such as H-II regions. Also includes an improved statistical noise matching engine that matches the noise statistics of surrounding areas when removing large stars, making these areas blend into the rest of the image.
7 16 Dec 2021 Revamped neural network architecture giving greater accuracy of star removal and better retention of non-stellar structures. Trained on an expanded data set that includes more galaxy images, resulting in better retention of galaxy cores and non-stellar details in spiral arms. AI7 retains much more detail around stars than AI5, largely correcting the overly-smooth areas of that version.
6 8 Nov 2021 Only available for PixInsight version at this time. Implements matching of noise statistics for replaced pixels, resulting in a more natural appearance and elimination of the "smooth" areas around stars seen in earlier AI versions.
5 26 Sep 2021 Improved handling of monochome narrowband and noisy nebula photos.
3 18 Sep 2021 Much better handling of wide-field (short focal length) galaxy images. Better handling of crowded fields of small stars from short focal length instruments. Better avoidance of small non-stellar features (e.g., small patches of nebulosity and background galaxies).
2 12 Sep 2021 Initial release

Is this StarNet++?

In case you are wondering if StarXTerminator is just the popular StarNet++ repackaged as a Photoshop plug-in or PixInsight module, it is most definitely not. The neural network architecture, the training data set, the training method, and the supporting algorithms for tiling and batch processing are completely different and unique.