NoiseXTerminator version 2, with AI version 3, is now available for installation in PixInsight 1.8.9-2 and later. The update will be distributed automatically using PixInsight’s update system and the NoiseXTerminator repository. This is a major update with changes listed below. An online NoiseXTerminator user manual is also now available.
Completely new neural network architecture and training method
The new architecture is much more effective at maintaining detail, even faint detail in dark regions, while also suppressing large-scale noise. In the past these two functions have been mutually exclusive – suppressing large-scale noise tended to also squash faint detail. AI3 decouples these in a very effective manner.
Iterative noise reduction
The new “iterations” parameter causes noise reduction to be done gradually, in the specified number of steps. Larger values tend to increase preservation of faint details while also suppressing faint low-frequency (large-scale) noise, at the expense of longer execution time. More iterations tend to be most effective on very noisy images. On extremely noisy images, a large number of iterations may cause subtle artifacts to appear – in this case, reduce the number of iterations.
Intensity/color noise separation
In this mode, selectable using a checkbox, NoiseXTerminator will apply different amounts of noise reduction to the intensity (brightness) and color noise in the image. Reducing the color noise more than the intensity noise can be very useful for maintaining a natural appearance.
High/low frequency noise separation
This mode allows different amounts of noise reduction to be applied to the high-frequency (fine-scale) and low-frequency (large-scale) parts of the noise. While AI3 is already much better at both preserving faint detail and suppressing large-scale noise, these adjustments allow further control over maintaining detail at certain scales. What’s more, the definitions of high- and low-frequency are adjustable using the HF/LF Scale parameter. This defines the spatial scale (in pixels) that separates the high and low noise frequency bands.
Combined modes
The color and frequency separation modes can both be used at once by selecting both checkboxes. This increases the number of parameters to adjust, but the degree of control this provides can be worth it. And one can always fall back to the “easy” mode of just specifying Denoise and Iterations.
Debayering and 2x drizzle
In contrast to AI2, AI3 comprehends artifacts produced by 2x drizzle and debayering. These are usually only seen when too few exposures have been taken and/or dithering was insufficient. AI2 can sometimes malfunction on these images, producing tiling and other odd artifacts.
Non-generative AI
When it comes to astrophotography, we want images that are as true to reality as possible. A distinguishing feature of all RC Astro tools – NoiseXTerminator, BlurXTerminator, and StarXTerminator – is that they are all trained in a non-generative fashion. As such they prioritize producing “maximum likelihood” results over the creation of false detail.
Photoshop
The Photoshop version of NoiseXTerminator 2/AI3 is coming soon!