PixInsight 1.9.5 Lockhart Update Brings Machine Learning and Speed

There is a fair amount of waiting in astrophotography, and some of it happens long after the telescope is put away. A folder full of exposures still has to become an image you can work with. PixInsight 1.9.5 Lockhart puts attention into that part of the job, with faster processing routines and a new noise reduction tool called MLDenoise. The update is available for Linux, macOS, and Windows.

What interests me here is how these changes fit into an ordinary processing session. Faster integration gives you more room to try a different rejection setting or revisit a stack. A useful noise reduction tool gives you another way to work with faint data. Both deserve a closer look, provided we keep the expectations tied to what the software actually does.

Less waiting for a stack

LocalNormalization, ImageIntegration, and DrizzleIntegration have all received performance work. Many of the mathematical routines underneath them have been rewritten to make better use of modern processors. That work reaches into operations used throughout PixInsight, including image statistics, interpolation, and measurements of star profiles.

The amount of time you save will depend on your computer, your data, and the settings you use. A large set of exposures can put very different demands on a machine than a small stack. I would judge the improvement with a familiar project, using the same files and settings, so the comparison tells you something useful about your own workflow.

The preview system also supports 32 bit floating point images in tools including MLDenoise and MultiscaleAdaptiveStretch. That matters when you are judging faint detail. The older preview could show banding that was absent from the finished result, which made it harder to decide whether an adjustment was helping.

Where the PixInsight 1.9.5 Lockhart update brings machine learning and speed into practice

MLDenoise is the first process in the new MachineLearning module. It uses neural networks for noise reduction, with an Amount control that blends the processed image with the original. There is also a Local Support option, which builds a mask from the image and lets you vary the treatment with brightness. You can inspect that mask and evaluate the result in the preview before applying it.

Those controls matter more to me than the machine learning label. A faint dust cloud and an uneven background can be difficult to separate, especially when the data is thin. Smoothing the image until it looks clean is easy enough. Keeping the small variations that belong to the object takes more judgment. I would compare the result with the original at a consistent scale and pay attention to the faint structures as well as the background.

The models are downloaded separately from the application and can be updated independently. Processing can use the CPU or supported GPU hardware, including NVIDIA CUDA and CoreML on macOS. That gives users several ways to run the tool, but the experience will still depend on the hardware available.

Noise reduction remains one step in processing. It cannot replace exposure time, calibration, or a well focused telescope. My interest is in whether it makes useful data easier to work with while leaving the image believable. That is a more worthwhile test than seeing how smooth the background can become.

Getting the image coordinates right

The astrometry changes are less obvious at first glance, but they are worth understanding if you work with mosaics or fields that show optical distortion. ImageSolver has a new option called Recursive surface splines. It models distortion locally across the image using matched stars, rather than relying on a limited selection of points for one global fit.

An astrometric solution connects positions in an image to coordinates on the sky. Improving that model can help when placing images by their coordinates, particularly across a large field. It does not guarantee an invisible mosaic seam. Background differences, gradients, and variations in the data still need attention.

PixInsight also stores the complete solution in a documented part of the XISF format, including its distortion model. The solution can be restored when the file is opened without rebuilding it. There is a compatibility detail to remember here: version 1.9.5 can read older solutions, but earlier versions cannot read the new solution format. That matters if you move projects between computers or share data with someone using an older installation.

A few things to do before processing

This release needs a manual installation of the core application. It does not arrive through the usual application update process. After installing it, run Thread Performance Analysis from the Process menu. The processing routines have changed enough that the old measurements are no longer suitable, and fresh measurements help PixInsight choose how many processor threads to use.

WBPP and FBPP now share version 3.1.0. Apple Silicon users also gain access to the legacy SpiderMonkey script engine in the native application, allowing older scripts that depend on it to run without using the Intel application through Rosetta. There is a qualification: the legacy engine runs as an interpreter on Apple Silicon, so native support does not mean every older script runs faster.

For me, the sensible way to evaluate an update like this is to return to a data set I already understand. Check how long the stack takes, compare the noise reduction carefully, and see whether the tools make decisions easier. Processing should help us understand and present what we collected. If this update saves time while giving us better control over that work, there is a practical reason to use it.