Parallax: sharpen and
round the stars
SyQon's aberration correction, star reduction, and non stellar sharpening. The working reference for running it in Seti Astro Suite Pro, PixInsight, Siril, and Photoshop.
Parallax is SyQon's neural sharpening engine, the third tool alongside Starless for star removal and DeepPrism for noise reduction. It does three jobs that older deconvolution tools struggle to do cleanly at once: it corrects optical aberrations in the stars, reduces star size, and sharpens the non stellar structure. The three jobs are separate models, so you can run them together or one at a time. It runs in Seti Astro Suite Pro, PixInsight, Siril, and Photoshop, with a free Nano tier and a paid Pro tier.
The three models
Parallax ships as three model files, each handling one task. Knowing which model does what tells you which slider to reach for. Any of the three can be switched off, so none is forced on you.
Aberration correction
Rounds off coma, elongation, and spherical aberration in the star field. A simple on or off toggle.
Star reduction
Shrinks star size. Zero is off. Default is three, with four to five a good working range.
Non stellar sharpen
Deconvolves nebulosity and galaxy detail. One equals 100 percent strength.
In correct only mode, Parallax and BlurXTerminator produce identical stars. Toggling between them on a test frame shows no difference at all. The two tools only diverge once you sharpen or reduce, where they take very different approaches. Rounding a star is rounding a star.
Getting it installed
Seti Astro Suite ProUpdate the suite, then load the models
Parallax needs version 1.18 or later. Grab it from setiastro.com under Astro Programs. Install routes: the standard Windows or macOS installer, a pip install on any platform, or Homebrew and the GitHub repo on macOS. Then click Get Models in the Parallax panel, buy, and unzip the three model files into the correct model folder.
PixInsightTwo parts: a CLI engine and a script
Each SyQon product is two pieces: a CLI engine (the zip from your syqon.eu dashboard) and a PixInsight script that drives it. The zip alone will not show up in PixInsight, so you need both.
- From the SyQon Download Center, download your builds and extract each zip to a permanent location. Keep every *_cli.exe with its _internal folder, since they do not work apart.
- In PixInsight open Resources › Updates › Manage Repositories and add the SyQon scripts repository: raw.githubusercontent.com/SyQon-Hub/PixInsight_Scripts/refs/heads/main/
- Open Resources › Updates › Check for Updates, apply, then restart PixInsight.
- Open each script and, in the Executable section at the bottom, point it at the matching engine: parallax_cli.exe for Parallax, prism_cli.exe for DeepPrism. The path is remembered.
Licensing is account based: it verifies against your SyQon login on first run, with no key to paste. The first DeepPrism run can be slow while its roughly 400 MB model loads. If you move the engine folders later, re-point the scripts. A manual install bypasses the update system, so pull new script versions from the SyQon GitHub repo when they are released.
SirilInstall the script, then the AI weights
Parallax, DeepPrism, and Starless (star removal) all run as Python scripts in Siril.
- Open Scripts › Get Scripts. Search SyQon, or the tool name (parallax, prism, starless). After a recent reshuffle they may briefly list under the Other category, a Windows display bug a future Siril release will fix. Select and hit Apply.
- Find them under Scripts › Python › SyQon.
- Open Parallax. Pick Nano or Pro at the top, then click Get Free Models or Get Models. A SyQon account is needed to download or buy.
- Download the three AI weights: aberration correction, star reduction, and stellar sharpening.
- Use the three Install buttons to load each weight from your downloads. They are copied to a fixed location, so the originals can then be deleted.
Running it on an image
- Prepare the frame first: crop, run background extraction (AutoBG or your preferred tool), plate solve via Tools › Geometry › Plate Solver, then colour calibrate with SPCC (untick Plot FITS). Save, so the prep is not lost.
- With the data still linear, open Scripts › Python › SyQon › Parallax.
- Choose Nano or Pro, then set the three models: aberration correction on or off, star reduction 0 to 10, sharpen strength. Drop any you do not want to zero.
- Leave Apply temporary stretch on for linear data, set target median to about 0.12 for a balanced preview, and turn Linked Stretch on if the stars are still in the image.
- Click Process. It takes a few minutes, so maximise the window to read the result.
- Review on the divider that splits the view, original on the left and processed on the right: scroll to zoom, hold left mouse to pan, drag the divider to compare. When happy, click Import in Siril to bring the result back into your image.
Photoshop is supported as well, using the same paid models. The current installer is on the SyQon website.
Free or Pro
Both tiers use the same controls and the same three models. The difference is how much model sits behind them. Nano is free with no usage limits and does a genuinely capable job. Pro is trained on far more and tends to edge ahead, especially on demanding data. Start free, upgrade only if you want the last bit of quality.
| Aspect | Nano · free | Pro · paid |
|---|---|---|
| Model parameters | 1.3 million | 4.5 million plus |
| Workers | single | multi-threaded, up to seven |
| Aberration correction | basic spatial | coordinate-aware advanced |
| Star reduction | model only | hybrid circular erosion |
| Platforms | Siril only at time of recording | all-in-one bundle, or per host |
Under the bonnet, Pro runs a deeper residual, coordinate-aware network (SyQon's StellarDirectNet Pro) with squeeze-and-excitation attention and multi-scale feature aggregation, tuned to converge up to ten times faster on consumer hardware.
Price depends on the host: €87 for PixInsight and Photoshop, €67 for Siril and Seti Astro Suite. A discount code gets 5 percent off in the cart. Choose the all-in-one bundle for every host, or pick a single host from the pulldown. After purchase, download from your account, extract the zip, ignore the MacOS folder, and install the three weights exactly as you would the free ones. Settings are identical across both tiers.
The controls
| Control | Range | What it does |
|---|---|---|
| Aberration correction | on / off | Toggles the correct only model. Rounds distorted stars across the frame. Identical result to BlurXTerminator's correct only. |
| Star reduction | 0 to 10 | Zero is off, default is three. Four or five is the sensible working range. Seven is already strong. Ten does odd things and is not advised. |
| Sharpen strength | 0 to 2.0+ | One is 100 percent. Zero skips the model. Around 120 percent pulls more structure. Past 2.0 the field turns red with a warning, since it oversharpens unless you have a strong blur to fight. |
| Tile size | 512 default | The AI works in tiles. Siril defaults to 512. Drop toward 128 on an older GPU such as a GTX 1650, or when VRAM is tight. |
| Overlap | 64 default | How much neighbouring tiles overlap to blend the seams. Must always be a smaller value than the tile size. |
| Edge pad | default | Mirrors the image's own border pixels outward, so reassembling the tiles does not leave a hard cut-off at the edge. |
| Target median | ~0.12 | The level of the temporary preview stretch, adjusted live. Aim for a balanced background, not too bright. Not critical, it only guides the model. |
| CPU execution | fallback | Forces processing onto the CPU if the GPU misbehaves. Slower, but it gets the job done. |
| Clear AI cache | rarely | Loaded models stay in memory between runs. This wipes them and forces a reload from disk. Troubleshooting only. |
A new Parallax release lifted star reduction from a maximum of six to ten, and sharpen from 100 to 200 percent. More usefully, the star reduction model was retuned: the old very soft behaviour is gone, halos are reduced, and the result now sits much closer to BlurXTerminator, though BXT still reduces a touch harder.
Linear stage and colour
Deconvolution belongs in the linear stage. It enhances detail that is almost invisible there; once you stretch, that detail is effectively gone and you are only sharpening what remains. This holds for Parallax and BlurXTerminator alike. Parallax will run on stretched data if you must, but linear is always the stronger choice.
If your image is still linear, leave the temporary stretch enabled so the model has something to work with. It is applied only for the model and does not bake into your data.
If the stars are still in the image, or you have already run SPCC on a one shot colour frame, turn Linked Stretch on. Without it the channels stretch independently and you lose star colour. Linked stretch keeps the colour balance intact.
Order of operations
Colour-calibrate first, then recover structure, then denoise. In Siril a typical run is crop, background extraction, plate solve, then SPCC, then Parallax. Both ecosystems follow the same logic: deconvolve before you denoise, so the denoise step cleans an image whose structure is already resolved.
SyQon's recommended starting point is Parallax then DeepPrism, which lets Parallax read the image before a denoise pass alters its texture.
SyQon are clear that this is a starting point, not a law. Both tools are neural and interpret whatever they receive, so the best order depends on the data. On a very noisy or faint image, a light DeepPrism pass before Parallax can steady the data so Parallax does not mistake noise for structure. Keep that first pass gentle, since a heavy denoise hands Parallax an image that has already lost fine texture. The main denoise still belongs after Parallax.
Compare a few versions: standard Parallax then DeepPrism, a lighter Parallax then DeepPrism, and a very light DeepPrism then Parallax then a second light DeepPrism. Judge at 1:1, not just the whole frame. Are the stars compact or too harsh? Are faint structures genuinely clearer, or starting to look reconstructed? Is the background clean, or going plastic? Has noise dropped naturally, or turned into odd texture? A little natural grain reads as more believable than a glassy sky.
Parallax plus DeepPrism for nebulae
The pairing is where the real gain shows, the same way BlurXTerminator pairs with NoiseXTerminator. Tested on the Orion Nebula it is not subtle: structure that was only hinted at comes out cleanly, and the field reads sharper without turning harsh. Do not push Parallax past what the data supports; if the stars go hard or artefacts appear, ease the strength back before the denoise pass.
DeepPrism is strongest on nebulosity. On galaxies and smaller objects, early versions could lay a strange pattern across the sky, and at default settings it has been reported to look blotchy on faint or noisy data. SyQon have treated this as a fault to fix rather than invented detail, so behaviour may differ by version. Lean on it for nebulae, and check it carefully on galaxy data before trusting the result.
Working through targets
Different targets ask for different model combinations. These are the settings used across the worked examples in the source walkthroughs.
Run correct only first to judge the aberration model on its own. Elongated and stretched corner stars pull back to round points, including the worst stars right in the far corners.
SCT corner stars get ugly when focus drifts, and the classic spherical aberration bloats them badly. Correct only cleans even heavily defocused corners and rescues fainter stars that other tools leave smeared.
Crop, background extraction, plate solve, and SPCC first, then Parallax with all three models on. The effect is deliberately subtle, which is what you want. Push past 2.0 and it goes crunchy. Corner aberration tightens and stars reduce slightly.
Used to test Parallax and BlurXTerminator head to head at matched strength. Both lift the structure, but at the same 100 percent Parallax pulls clearly more detail, with little black lanes appearing that BXT left flat.
Everything reads sharper with more contrast across the frame. Faint structure lifts out, down to small Bok globules, and the result stays smooth rather than picking up dimpling in the shadows.
The showcase for the SyQon pairing. Parallax brings the structure out, DeepPrism cleans behind it. Against BXT plus NXT the gap is striking: far more visible structure and a much sharper, cleaner field.
The stars are already pinpoint round, so Parallax barely touches them beyond clearing chromatic aberration. The gain is in the galaxy itself, where it pulls out real dust and structure that BXT leaves softer.
Dust lanes pop and the core resolves into small knots and clusters, which is correct for a galaxy. No long wormy artefacts. DeepPrism needs a careful eye on galaxy data, so check it rather than trusting it blind.
At 100 percent the dark detail, bright contrast regions, and crepuscular ray structure all come up cleanly. No need to push to 200 percent here.
Parallax vs the alternatives
Deconvolution reverses the blurring that the atmosphere and optics introduce. It recovers real detail and tightens stars by correcting aberrations such as coma and astigmatism, and done well it is the most transformative sharpening step in deep sky processing. The headline question is usually Parallax or BlurXTerminator, but that framing misses the point: the two tools are built on different ideas, and the right one depends on the kind of astrophotographer you are.
Two philosophies: the stars
The first time you run Parallax the stars can look like nothing happened. That is the design, not a fault. The two tools treat star flux in opposite ways.
- Boosts the flux, so stars become bright and tight, often pure white.
- You can shrink them as far as you like.
- The trade is that this is no longer their real brightness, and often no longer their real colour.
- Removes the variation to make the star round while keeping its real brightness.
- Holds star colour and clears chromatic aberration.
- Measurably rounder than BXT, but softer and not as small, since the flux is never crushed. If you want them smaller, a dedicated star reduction tool does that afterwards.
There is a long running debate that BlurXTerminator can sometimes invent detail. Parallax takes a different route by design. SyQon describe it as a physics compliant, gated, coordinate-aware residual regression network of around 1.3 million parameters: it maps and corrects spatially variant distortions such as coma, astigmatism, and structural flexure while extracting real fine structure, with a non-generative residual design that applies a bounded correction rather than painting in guesses. When the model is not confident it leaves the data alone, and 16 and 32-bit float FITS are preserved without quantisation loss. It is conservative on purpose, so that everything it does is real.
Against Cosmic Clarity
In the Seti Astro Suite Pro world, the free comparison is Cosmic Clarity.
Parallax
- More mature correct only model. Stars come back rounder, faint stars handled better.
- Sharp and smooth output with no dimpling in the shadows.
- Pulls more genuine structure from the deconvolution pass.
- Galaxy detail resolves to knots and clusters, not worms.
Cosmic Clarity
- The best free deconvolution and sharpening tool available.
- Corrects aberrations well, though heavy elongation can linger.
- Prone to dimpling in shadow regions, typical of deconvolution.
- Adds noise structure that wants a denoise pass afterwards.
Against BlurXTerminator: structure
Set the star treatment aside and the structural story flips. On nebulosity and galaxies Parallax pulls far more real detail, and BXT can look soft by comparison. On the Elephant Trunk and Orion the difference is dramatic.
Parallax
- Far more real detail on nebulosity and galaxy structure.
- That detail is genuine, not hallucinated, by design.
- Stronger still when paired with DeepPrism on nebulae.
BlurXTerminator
- Can look soft on structure next to Parallax.
- Tighter, sharper stars and the look many still prefer.
- Considerably faster.
For a fair test the source matched settings: Parallax at star reduction 4 and sharpen 80 percent, against BXT at 70 percent sharpen, 40 percent star reduction, and halo removal off, since Parallax does not do halo removal.
In correct only mode, yes. Tested on Bode's Galaxy, correct only on each tool is 100 percent identical. They only differ once you sharpen or reduce.
No difference. Star removal works the same whether you ran Parallax or BlurXTerminator beforehand. The visible difference between the two frames is the nebulosity, where Parallax looks better defined.
Overstated. Croman's own documentation says it does not, and pixel checks bear that out. Only a handful of the very brightest stars sit near clipping. Better understood through the flux story above: BXT whitens by boosting flux, Parallax keeps the real values. Two valid looks, your taste decides.
On the same frame, roughly three times slower than the exterminators. Still under a minute, so not a real obstacle.
Who is it for
A clear yes
Siril, Seti Astro Suite, and Photoshop users finally get strong deconvolution where there was none. It levels the field.
Buy without fear
If you avoided BlurXTerminator over hallucination worries, Parallax is built not to invent detail and stays scientifically accurate.
Your call
It comes down to whether you prefer natural colourful stars, and whether the extra real structure is worth the cost.
On high focal length targets where you do want smaller stars, run BlurXTerminator for the stars with its non stellar sharpening switched off, then run Parallax with star work off and focused only on the structure. Stars the BXT way, structure the Parallax way.