GraXpert's AI noise reduction, run from DSPX
GraXpert Denoise reduces the noise of an image with the AI model of GraXpert, a separate program that we run for us. We choose the strength, DSPX sends the image to GraXpert and brings the result back as the new image. GraXpert is not part of DSPX: we install it ourselves.
GraXpert is free software for astrophotography, made by its own authors and available from its own website. Its denoising uses a neural network that works on the image in tiles. This process takes the image of the panel, hands it to the GraXpert installed on this computer, waits for it, and replaces the image with what GraXpert made. The result goes through the image's mask and selection and makes one history step, like the result of any other process.
DSPX ships nothing of GraXpert and downloads nothing of it. GraXpert itself is free software under the GPL-3.0, and the AI models it uses have their own license, Creative Commons BY-NC-SA 4.0, which is for non-commercial use. They are separate software, and their licenses are theirs. We say which program to run in . Our own noise reduction, which needs no neural network, is Noise Reduction.
The panel has two sections, Denoise and Model, then the image the process applies to, the buttons and the system bar along the bottom. There is no live preview: GraXpert is slow, so the process runs when we click Apply.
Strength: How much noise is removed, from 0 to 1. 0.5 by default, as in GraXpert.
Batch size: How many image tiles GraXpert processes at once, from 1 to 32. 4 by default. A larger batch is faster and needs more memory, on the graphics card or the processor.
Model version: The version of GraXpert's AI model, written as three numbers such as 1.0.0. Empty, the default, uses the latest one GraXpert has. A model that GraXpert does not have on this computer yet is downloaded by GraXpert itself.
Apply writes the image to a temporary file, runs GraXpert on it, reads the result and replaces the image with it, in one history step that Undo takes back, through the image's mask if it has one. The panel shows the progress GraXpert reports, and the Cancel button stops GraXpert and everything it started. The image keeps its size, its channels and its sample type; integer images get the result clipped to their range, so nothing wraps around. Alpha channels and the image's metadata, astrometric solution included, are not touched.
We want to be exact about what DSPX does on this computer.
-cli -cmd denoising, the
file, -output result, then -strength, -batch_size, -ai_version when we
name a model and -gpu when GPU in Preferences is On or Off. The first line DSPX writes to the Log is
this command.Compare with Undo. Apply, look, undo and apply again with another Strength. The History panel keeps each step.
Out of memory. A smaller Batch size uses less memory.
Only part of the image. A selection or a mask limits the result to its area, as for every process; GraXpert still works on the whole image, so the run takes as long.
Noise Reduction: our own noise reduction, without a neural network, with a live preview.
GraXpert Background Extraction: GraXpert's gradient removal.
GraXpert Deconvolution: GraXpert's deconvolution, in the releases that have it.