Stretch

Convert linear data -- where almost all real signal sits near zero and is invisible on a normal display -- into a non-linear image a human can actually see, without crushing the sky, clipping stars, or hiding the shape of the stretch curve's own tradeoffs.

schema 1 · revision 1

Recognize it

The linear image displays as almost entirely black except star cores; faint structure is present in the data but not visible without a tone curve.
A naive linear stretch either leaves the image too dark to judge or blows out the brightest cores before faint structure becomes visible.

Intended output

A non-linear image with a dark, low sky background, visible faint structure (galaxy arms, dust lanes, nebulosity), and star cores that are not hard-clipped.

Required input state

Linear, unstretched data, after deconvolution and denoise (and star correction, on workflows that run it) -- late in the pipeline, immediately before background-neutralize/color-boost operate on the now-visible tonal range.

What NOVA does

NOVA's ontology exposes six real stretch engines behind 13 selectable parameter presets: PixInsight MultiscaleAdaptiveStretch (pi_mas -- M66's recorded path), SASpro Statistical Stretch (stat_default/stat_bright/stat_globular), SASpro GHS Stretch (ghs_default/ghs_galaxy), and three NOVA-original engines with no external tool behind them: stf_stretch (a from-scratch reimplementation of the STF tone-curve concept, presets stf_galaxy/stf_nebula), veralux_stretch (an original arcsinh-based design with auto-detected symmetry point, presets veralux_default/colorful/globular), and smart_stretch, which profiles the image's dynamic range and star fraction and adaptively selects among STF/veralux/GHS/stat rather than running one fixed engine.

Use when

Always, once linear processing (calibration, background, color, deconvolution, denoise) is complete -- this is the step that makes the image viewable at all.

Skip when

The input is already non-linear (re-stretching a stretched image compounds tone curves and is a distinct, deliberate operation, not this step).

Limits

This is the single most contested, aesthetically-loaded step in the whole pipeline -- 'correct' output depends on target type, sky darkness, and taste, not just physics. Treat any single stretch as a judgment call, not a ground truth.
M66's recorded run validates exactly one engine (PixInsight MultiscaleAdaptiveStretch). NOVA has six real, code-confirmed stretch engines total; the other five are real alternatives it can select for other targets, not M66-validated.
Different engines are mathematically unrelated -- statistical (percentile-based), generalized hyperbolic, arcsinh, and multiscale-adaptive curves. A similar-sounding parameter value on two engines (e.g. a 'target' near 0.08-0.15) does not imply a similar result; only the engine and its own documented math define the curve.
stf_stretch, veralux_stretch, and the adaptive selector smart_stretch are NOVA-original code with no `setiastro` import and no PixInsight or SASpro GUI equivalent. veralux_stretch independently reimplements a real, separately installed Siril Python script named Veralux -- prefer that real script (Siril tab) as the reference path; Siril's native Asinh Stretch remains a fallback approximation when the actual script isn't installed. stf_stretch and smart_stretch still have no manual path in any tool.
Irreversible: clipped highlights and crushed shadows from an aggressive stretch cannot be recovered from the output alone; the input's actual dynamic range sets the ceiling on what any engine can safely show.

Scientific and aesthetic notes

Jeff's own stated preference is galaxies darker and higher-contrast than a generic 'bright and colorful' default -- a stretch that looks technically correct can still be the wrong aesthetic call for a given target.
smart_stretch's adaptive engine selection is itself a real NOVA design decision, not a documentation gap: for some targets the 'right' engine genuinely depends on measured dynamic range and star density, not a fixed per-target-type default.
The historical MAS parameters recorded from M66 are the module's own installed defaults (version 1.1.1.0), not values NOVA computed or tuned for this image -- the workflow created the process and ran it with zero overrides.

Evidence and measurements

beforeafter

stretch before and after

Cross-tool matrix

Tool	Equivalence	Expected result	Evidence
nova	Exact replay	A non-linear image with sky near-black, visible faint structure, and unclipped star cores.	nova-stretch-source, nova-m66-run-1.24.7
pixinsight	Exact replay	Full-array median ~0.085, p99 ~0.473, max ~0.936 on M66's linear starless input -- deterministic (byte-identical output across repeated runs on the same input).	m66-manual-verification, nova-stretch-source
siril	Same engine, adapted host	A statistically-targeted stretch with the sky near the target median and negligible hard clipping.	m66-manual-verification
siril	Algorithmically equivalent	An arcsinh-based stretch from the same named script family NOVA's veralux_stretch reimplements, closer to NOVA's engine than the native Asinh Stretch fallback.	jeff-siril-veralux-note
siril	Functional alternative	A workable stretch using Siril's own native tools; not a reproduction of any specific NOVA engine's exact curve.	siril-native-stretch-manual
siril	Conceptual substitute	A dark-sky, arcsinh-shaped stretch broadly similar in character to Veralux, without matching its exact curve, color handling, or automatic background detection.	siril-native-stretch-manual
saspro	Same engine, adapted host	Sky lands near the chosen target median with linked-channel color preserved and a mild curves-boost S-curve applied.	saspro-source-1.18.0, nova-stretch-source
saspro	Same engine, adapted host	A hyperbolic-curve stretch with the pivot anchored near the sky level.	saspro-source-1.18.0, nova-stretch-source

NOVA Python pipeline workflow 1.23.0 (M66 run); ontology current

Confirm which of the six engines the active workflow selects before assuming MAS (M66's engine) is the general default -- it is one of six, not the only one.
For stf_stretch or smart_stretch, do not look for a manual reproduction path in PixInsight/Siril/SASpro; none exists for these NOVA-original engines. Judge the output on its own measurements instead.
For veralux_stretch, prefer the real Veralux Siril script (see the Siril tab) -- the same named algorithm, independently reimplemented -- over the native Asinh Stretch fallback if the actual script is installed.
For stat_stretch/ghs_stretch/mas, the tool-specific tabs below give the real manual reproduction path.

Controls and starting ranges

M66 recorded engine: pi_mas -- PixInsight MultiscaleAdaptiveStretch, module defaults, no override
other selectable engines: stat_stretch, ghs_stretch, stf_stretch, veralux_stretch, smart_stretch
NOVA-original, no manual equivalent anywhere: stf_stretch, smart_stretch
NOVA-original, real Siril script counterpart: veralux_stretch -- Siril's own Veralux script, algorithmically equivalent; native Asinh Stretch is only the fallback

Mask behavior: Engine-specific; MAS has an internal lightness mask on its color-saturation stage (see the PixInsight tab).

Failure modes

Presenting M66's MAS result as NOVA's one and only stretch behavior.
Assuming stf_stretch or smart_stretch has a PixInsight/Siril/SASpro equivalent.
Reaching for Siril's native Asinh Stretch fallback when the real Veralux script is actually installed and available.
Presenting NOVA's veralux_stretch and the Siril Veralux script as byte-identical just because they share a name and algorithm -- independent implementations, algorithmically equivalent, not exact replay.

Recovery

Revert to the linear input and select a different engine/preset for the target's actual dynamic range.

PixInsight MultiscaleAdaptiveStretch (MAS) MultiscaleAdaptiveStretch module 1.1.1.0

Create a MultiscaleAdaptiveStretch instance and run it on the full linear starless image -- not a partial preview; Contrast Recovery does not permit partial-preview execution.
M66's run applied every module default with zero parameter overrides; only override noiseThreshold/clippingFraction if you have a specific reason to.
Inspect the whole-image statistics after; they describe the whole array, not a claim that the empty-sky region specifically landed at the target background.

Controls and starting ranges

Target Background: 0.150 (module default)
Aggressiveness: 0.70 (module default)
Dynamic range compression: 0.40 (module default)
Background Reference: disabled (module default)
Contrast Recovery: enabled, scale separation 1024px, intensity 1.00 (module defaults)
Color Saturation: enabled, Amount 0.75, Boost 0.50, Lightness mask enabled (module defaults)

Mask behavior: MAS's own internal Lightness mask restricts its color-saturation stage by luminance; this is not an externally-applied PixInsight mask.

Failure modes

Substituting manually-tuned values (e.g. Target Background 0.095, Aggressiveness 0.80) and calling the result an exact NOVA reproduction -- those are Jeff's own experimental settings, not what the pipeline ran.
Running Contrast Recovery on a partial preview, which the module does not support.

Recovery

Reset to module defaults and re-run on the full starless image.

Seti Astro Statistical Stretch via Siril's Python integration Seti Astro Statistical Stretch integration v3.1.2, verified in Siril 1.4.4

Open the licensed Seti Astro Statistical Stretch script on the linear image.
Set the target median and black-point sigma, enable Linked Stretch and Normalize.
Apply the curves-boost pass at the recorded strength; check the reported clip percentage before accepting.

Controls and starting ranges

M66 verified: Target median 0.15; Black point sigma 5.00; No black clipping, Linked Stretch, Normalize all enabled; HDR Highlight Compress disabled; Stretch luminance only disabled; Apply curves boost enabled at strength 0.50
M66 result: preview clip report 1 of 1,352,337 pixels (0.0001%)

Mask behavior: Not demonstrated; operates on the full image.

Failure modes

Treating this as a stock Siril feature -- it is a separately licensed external integration, not bundled with Siril.

Recovery

Undo and re-run with a lower target median or higher black-point sigma if the sky lifts too far.

Veralux via Siril's Python integration separately installed Python script; exact version not independently confirmed in this repo

Open the Veralux script from Siril's Python scripts on the linear image -- a separately installed script, not a stock Siril feature, the same pattern as the Seti Astro Statistical Stretch integration above.
This is the real, named script NOVA's veralux_stretch independently reimplements in Python -- prefer this over the native Asinh Stretch fallback below when it's installed, since it shares the actual algorithm rather than only the general arcsinh math.
Record the exact menu path, script version, and every setting used; none of that is independently confirmed in this repository yet.

Controls and starting ranges

availability: Jeff-confirmed installed and available; exact menu path, version, and default parameters not yet independently verified in this repo

Mask behavior: Not independently confirmed in this repo.

Failure modes

Treating this as a stock Siril feature rather than a separately installed script.
Assuming NOVA's independently-coded veralux_stretch reproduces this script byte-for-byte just because they share a name and algorithm family -- same-named algorithm, independent implementations.

Recovery

Fall back to native Asinh Stretch (below) if the script isn't installed.

Native Generalized Hyperbolic Stretch / Histogram Transformation 1.4.4

Use as a no-extra-cost native alternative when the Seti Astro integration isn't installed.
For GHS: set the Symmetry point just above the background, raise D (stretch intensity) and b (local intensity) to lift signal without blowing the core.
Autostretch (the eyeball icon) is a fast preview, not a final result.

Controls and starting ranges

GHS starting point: Symmetry point near background level; no fixed D/b range established

Mask behavior: Not demonstrated.

Failure modes

Presenting a native GHS/HT result as equivalent to NOVA's MAS, Statistical Stretch, or any NOVA-original engine.

Recovery

Undo and adjust the symmetry point or stretch intensity.

Native Asinh Stretch (fallback if the Veralux script isn't installed) 1.4.4

Use only as a no-extra-cost fallback when the actual Veralux script (above) isn't installed -- both are arcsinh-based, but this is Siril's own separate native tool, a conceptual approximation rather than the same algorithm.
Open Image Processing > Asinh Stretch on the linear image; its two controls are stretch factor and black point.
Sample the empty sky first and set the black point from that measurement, mirroring Veralux's own auto-detected symmetry point, before raising the stretch factor.
Increase the stretch factor gradually and check star cores and color saturation at each step; Siril's default RGB blend clipping mode does its own luminance-based color blending during the stretch, but it has no adjustable equivalent to Veralux's color_grip parameter or its shadow-convergence damping.

Controls and starting ranges

black point: set from a measured empty-sky sample, not left at a default; no fixed value established
stretch factor: no fixed value established; raise gradually and inspect
not reproduced here: Veralux's adjustable color_grip parameter and shadow-convergence damping have no Asinh Stretch equivalent; Siril's own default RGB blend clipping mode blends luminance and color differently and isn't a substitute

Mask behavior: Not demonstrated.

Failure modes

Leaving the black point at Siril's default instead of sampling the actual sky level.
Presenting this result as a reproduction of Veralux rather than an approximation of the same general math.

Recovery

Undo and resample the black point, or reduce the stretch factor if color saturates unevenly.

Statistical Stretch 1.18.0 source-inspected

Run Statistical Stretch on the linear image with Linked channels enabled.
Choose a preset by target: default for an ordinary galaxy/broadband target, brighter for a dim target or very dark sky, globular for a dense star cluster.
This is the same underlying algorithm Siril's licensed Seti Astro integration exposes -- both trace to the same statistical-stretch design, hosted differently.

Controls and starting ranges

stat_default: target_median 0.08, linked, blackpoint_sigma 4.0, curves_boost 0.05 -- sky should land ~0.06-0.08
stat_bright: target_median 0.13, linked, blackpoint_sigma 4.0, curves_boost 0.05 -- dim targets or very dark skies
stat_globular: target_median 0.11, linked, blackpoint_sigma 5.0, curves_boost 0.05 -- Jeff-confirmed good on C 80 (8.2/10 vs a pipeline stf_nebula run's 6.2/10); lowered from 0.15 after a globular over-stretched with core clipping

Mask behavior: Not applicable; full-image linked stretch.

Failure modes

Using the galaxy/broadband preset on a dense globular, which needs the extra shadow headroom the globular preset provides.

Recovery

Revert to the linear input and choose the preset matching the actual target density.

GHS Stretch 1.18.0 source-inspected

Run GHS Stretch on the linear image with the pivot set near the sky background, not a generic 0.25.
In this pipeline's automatic parameter selection, only alpha (stretch strength) and pivot are data-driven from the image's dynamic range; beta, gamma, lp, and hp stay at their function defaults regardless of target -- do not assume every GHS control is being tuned per-image.
Choose the galaxy preset for a stronger core-to-arm range; the default preset otherwise.

Controls and starting ranges

ghs_default: alpha 5.0, beta 0.0, gamma 3.0, pivot 0.02, channel K
ghs_galaxy: alpha 8.0, beta -0.2, gamma 4.0, pivot 0.02, channel K -- stronger core-to-arm range
pipeline auto-selection: alpha and pivot only; beta/gamma/lp/hp are not varied by this pipeline's own parameter logic

Mask behavior: Not applicable; full-image operation.

Failure modes

Setting the pivot well above the sky background, which is a documented way to lift real signal into false brightness rather than stretch from the true floor.

Recovery

Revert and re-anchor the pivot to a measured background level.

Validation status

Not yet Jeff-validated. Use this as sourced guidance, not a certification.

Sources

M66 manual recipe verification feedback M66_Manual_Recipe_Verification_Feedback.md — Jeff validated
NOVA stretch engines: stat_stretch/ghs_stretch (real SASpro calls), stf_stretch/veralux_stretch/smart_stretch (NOVA-original, no external tool), PixInsight MultiscaleAdaptiveStretch, and ontology variants nas_server/seti_astro.py:273 (stf_stretch, pure NumPy, no setiastro import), nas_server/seti_astro.py:376 (stat_stretch, calls setiastro.saspro.imageops.stretch), nas_server/seti_astro.py:421 (ghs_stretch, calls setiastro.saspro.ghs_preset), nas_server/seti_astro.py:1136 (veralux_stretch, pure NumPy, NOVA-original, no setiastro import), nas_server/seti_astro.py:1319 (smart_stretch, adaptive orchestration over the above); nas_server/pixinsight.py:223 (mas flag); nas_server/pi_postprocess.js:813-833 (MultiscaleAdaptiveStretch PJSR call, no parameter override in M66's run); nas_server/tool_params.py:250-274 (compute_ghs: only alpha and pivot are data-driven, beta/gamma/lp/hp stay at function defaults); nas_server/processing_ontology.json:956 (13 experiment_variants across 6 engines) — NOVA source-confirmed
M66 workflow 1.24.7 execution record critiques/20260802_045251_seestar_galaxy.json — NOVA execution record
Project-authored Siril native-stretch guidance (SIRIL_MANUAL['stretch']) nas_server/recipe_page.py:116 — Reasoned translation
Jeff: Siril has a separately installed Veralux Python script (2026-08-15) Jeff, direct confirmation, 2026-08-15 conversation — Jeff validated
Installed Seti Astro Suite Pro 1.18.0 processing source saspro/pedestal.py; saspro/numba_utils.py; saspro/stacking_suite.py; saspro/abe.py; saspro/sssc.py — Artifact-confirmed

Related evidence

The M66 recipe and Jeff-validation ledger will link here when their static exports are available. Until then, this article remains explicitly unvalidated.