Curves

Refine where already-visible tonal zones sit and how strongly they separate -- shadow/midtone/highlight placement -- after the main stretch has already made the faint signal visible. Curves are final or near-final tonal placement, not a substitute for the main stretch and not a way to recover clipped or never-captured signal.

schema 1 · revision 1

Recognize it

The broad stretch is already acceptable, but the sky/background sits slightly too high or too low, midrange structure needs modest additional separation, or a bright core needs a smoother shoulder.
Near-white highlights lack headroom, or faint outer structure needs protection from an overly strong shadow pull.

Intended output

Intended tonal zones move in the intended direction while faint astrophysical structure that should not be sacrificed is preserved, with no new black/white clipping, no unintended non-monotonic tonal behavior, and color relationships preserved when the curve is intended as a linked/luminance-like operation.

Required input state

Nonlinear/stretched data with the upstream stretch method and effective parameters known, the relevant branch state known (current ontology expects the starless stretched branch before star recombination for the ordinary deep-sky path), and background/gradient/color-calibration problems already resolved rather than being disguised with tone shaping. Clipping/headroom and histogram percentiles should be measured before curves, and any luminance mask used should be recorded. A fair comparison requires the exact same parent nonlinear FITS for every candidate.

What NOVA does

seti_astro.curves() has two materially different execution paths. The SASpro preset/LUT path imports SASpro's curves_preset LUT machinery, exposing shape (linear/s_mild/s_med/s_strong/lift_shadows/crush_shadows/fade_blacks/rolloff_highlights), amount, and channel (all/R/G/B/L); it builds normalized control points into a 65,536-entry LUT applied to the (already globally min/max-normalized) data. The PixInsight path routes pi_curves/curves_pi to CurvesTransformation.K from named built-in shapes (including the two NOVA-authored globular strategies) or explicit custom control points. Separately, tool_params.compute_curves() derives measured control points from sigma-clipped sky background, percentiles (p80/p90/p95/p99/p99.5/p99.9-like), and object type, with object-specific sky-anchor priors and shadow-pull gentleness (gentler for galaxies than clusters) -- but as noted in limits, this computed strategy is not currently propagated into Experiment Mode's generic curve variants.

Use when

The broad stretch is already acceptable but one or more tonal zones need refinement -- sky level, faint-structure protection, midrange separation, highlight headroom, or object-specific core/halo balance.

Skip when

The image already has acceptable tonal placement.
The actual problem is a residual gradient, incorrect color calibration, noise, star halos, clipped acquisition data, or a fundamentally wrong initial stretch -- curves cannot fix any of these and should not be used to disguise them.

Limits

A tone curve is a pointwise nonlinear mapping (output = f(input)); it can redistribute recorded levels but cannot recover signal that was clipped, saturated, quantized away, or never captured. A lower p99 after a highlight rolloff demonstrates compression, not recovered highlight information.
NOVA's own data-driven strategy (tool_params.compute_curves(), deriving bespoke monotonic control points from measured sky level and percentiles) is computed inside Experiment Mode's adaptation path but its custom points are not currently propagated into the curve variants -- the generic adaptation loop only looks for a primary parameter called `amount`, and compute_curves() returns no `amount`. The current Experiment Mode curve run should not be described as testing NOVA's bespoke data-driven curve; that curve is computed but effectively discarded by this adaptation path.
Because compute_curves() supplies no `amount`, adaptation clamps every candidate's effective amount to [0.2, 0.8] using the nominal 0.5 default and each ontology ratio. This collapses `feather` (nominal 0.12) and `s_mild` (nominal 0.20) into the identical effective amount 0.20 with the same shape -- they are the same effective treatment and must not be counted as two independent candidates or two pieces of evidence.
The current SASpro-vs-PixInsight comparison is confounded by mask asymmetry: the luminance-mask map uses the key `curves_pi`, while the ontology's PixInsight candidate function name is `pi_curves`. Mask injection checks the exact function-name key, so SASpro `curves` candidates receive a computed midtone mask while `pi_curves` candidates do not, through the current default adaptation path. The current comparison is closer to masked-SASpro-strategy vs. unmasked-PixInsight-strategy than a clean engine comparison.
SASpro preprocessing is part of the effective treatment: NOVA's shared `_load_fits()` performs a global min/max normalization to [0,1] before the LUT is built, so a cross-engine comparison assuming bit-identical absolute sample values between engines is unsafe.
The `L` (luminance) channel mode explicitly quantizes through 8-bit OpenCV Lab before returning to float -- it is a distinct implementation strategy, not a high-precision floating-point luminance-curve equivalent, and should not be presented as one.
The two globular-specific PixInsight presets (`globular_balanced`, `globular_core_rolloff`) carry real, contradictory execution history: positive calibration comments from C80/Omega Centauri and M13, but later real runs (M12, M3) where the core-rolloff curve crushed halo/outer tones or sky background without the intended core benefit -- in the M12 case, an inversion guard correctly rejected the curve before it shipped. Two examples are not statistical validation, but they mean measured input state should outrank the object-class label alone for these presets.
The ontology's declared `none` candidate is not currently a working no-op control: the dispatcher only recognizes a copy-through when `engine` is explicitly `"none"` or `fn` is Python `None`, but the candidate sets `fn: "none"` as a literal string, so it falls through to `seti_astro.none`, which does not exist, and the run returns `ok: False`. Any comparison presented as validated against a real unchanged baseline is currently unsupported until this dispatch gap is fixed (tracked separately, not a content change).

Scientific and aesthetic notes

There is no single physically correct finished histogram for an astrophotograph. Curves have measurable constraints and failure conditions, but much of the final tonal preference remains aesthetic -- label perceptual judgments (flat vs. harsh, preferred sky darkness, visual emphasis) as rendering preference, not physical validation.
A rigorous comparison should separate distinct questions -- engine (match transfer function and mask as closely as possible), strength (same engine/shape/mask/parent, vary amount only), shape (same engine and strength/mask, vary shape only), mask strategy (same transfer function, full-frame vs. fixed mask), object-specific strategy (deliberately different curves/masks, classified as a strategy comparison), and adaptive-vs-fixed (the measured custom strategy against fixed presets and a no-op, once it is actually executed) -- rather than mixing all of them into one run's evidence.
Linked RGB curves (channel="all") apply the same function to R, G, and B, but f(R)/f(G) != R/G in general for nonlinear f -- linked curves can still change saturation/chroma relationships as a function of brightness and should not be described as exact color preservation merely because the same function was used per channel.

Evidence and measurements

Measurements	Acceptance criteria
Input/output control points or sampled transfer function, and its monotonicity. Background/sky median movement in fixed sky ROIs. Percentile movement (p05/p50/p80/p90/p95/p99/p99.9) on the exact same parent. Fraction at/near numerical 0 and 1. Local target-to-background contrast and radial/line profiles through a core, halo, arm, or filament, in fixed ROIs. Channel-specific clipping and color-ratio shifts. Difference images and protected-region residuals. Effective mask coverage and transition behavior, when a mask is used.	Intended tonal zones move in the intended direction without new black/white clipping or unintended non-monotonic behavior. Faint astrophysical structure not meant to be sacrificed is preserved, checked against a copied-input baseline on the same parent -- currently a manual copy rather than the ontology's own declared `none` candidate, which does not yet dispatch. Effective shape/control points, effective amount, channel mode, and mask state are recorded -- candidate ID alone is not sufficient, given the feather/s_mild collapse and the SASpro/PixInsight mask asymmetry. A lower highlight percentile is never presented as recovered information -- only as compression. Object-specific presets (especially the globular strategies) are checked against measured input state, not applied on class label alone, given the real contradictory M12/M3/C80 evidence.

Measurements

Acceptance criteria

Input/output control points or sampled transfer function, and its monotonicity.
Background/sky median movement in fixed sky ROIs.
Percentile movement (p05/p50/p80/p90/p95/p99/p99.9) on the exact same parent.
Fraction at/near numerical 0 and 1.
Local target-to-background contrast and radial/line profiles through a core, halo, arm, or filament, in fixed ROIs.
Channel-specific clipping and color-ratio shifts.
Difference images and protected-region residuals.
Effective mask coverage and transition behavior, when a mask is used.

Intended tonal zones move in the intended direction without new black/white clipping or unintended non-monotonic behavior.
Faint astrophysical structure not meant to be sacrificed is preserved, checked against a copied-input baseline on the same parent -- currently a manual copy rather than the ontology's own declared `none` candidate, which does not yet dispatch.
Effective shape/control points, effective amount, channel mode, and mask state are recorded -- candidate ID alone is not sufficient, given the feather/s_mild collapse and the SASpro/PixInsight mask asymmetry.
A lower highlight percentile is never presented as recovered information -- only as compression.
Object-specific presets (especially the globular strategies) are checked against measured input state, not applied on class label alone, given the real contradictory M12/M3/C80 evidence.

Cross-tool matrix

Tool	Equivalence	Expected result	Evidence
nova	Exact replay	Intended tonal zones move as intended with no new clipping, no unintended inversion, and protected regions stable.	nova-curves-source, nova-m12-m3-globular-curve-evidence
pixinsight	Same engine, adapted host	A monotonic tonal remapping via CurvesTransformation matching the requested shape or custom control points.	pixinsight-linear-nonlinear-p27, nova-curves-source
saspro	Same engine, adapted host	A 65,536-entry LUT tonal remapping applied to the normalized data.	nova-curves-source
siril	Functional alternative	A monotonic tonal remapping via Siril's own curve/histogram tools.	nova-curves-source

NOVA Python pipeline SASpro LUT path + PixInsight CurvesTransformation dispatch; measured compute_curves() strategy not yet propagated into Experiment Mode

Record the effective shape/control points, effective amount, channel mode, and whether a luminance mask actually applied -- candidate ID alone is not sufficient provenance, especially since feather/s_mild currently collapse to the same effective treatment.
Do not compare SASpro and PixInsight curve candidates as a clean engine test until the curves_pi/pi_curves mask-key mismatch is accounted for -- today one side is masked and the other is not.
Do not cite a historical curve-family win as validating NOVA's measured/adaptive strategy specifically; that strategy is not what the generic Experiment Mode variants currently execute.

Controls and starting ranges

s_mild / s_med / s_strong: SASpro LUT shapes, nominal amount 0.20/0.28/0.38
lift_shadows / rolloff_highlights: alternative tonal strategies, nominal amount 0.18/0.15
pi_s_mild_globular / pi_globular_core_rolloff: NOVA-authored PixInsight object-specific strategies; core_rolloff has contradictory real-run evidence
none: declared control, currently non-dispatching -- see limits

Mask behavior: SASpro path supports a computed luminance midtone mask via the generic nonlinear adaptation; PixInsight path does not currently receive it due to the curves_pi/pi_curves key mismatch.

Failure modes

Crushed shadows / faint-signal loss from an over-aggressive shadow pull.
Highlight inversion where a nominal rolloff brightens the core while lower tones darken (real M12 evidence).
Sky crushed without core benefit (real M3 evidence).

Recovery

Revert to the parent input; derive the shoulder from actual measured p95/p99; use a mask restricted to the intended region; weaken strength.

PixInsight CurvesTransformation CurvesTransformation.K; project runtime pin 1.9.3 Lockhart

A non-straight CurvesTransformation makes a previously linear image nonlinear -- apply after the primary stretch, not among physically linear operations.
Treat the two NOVA-authored globular presets as project strategies executed by PixInsight, not vendor-endorsed recommendations, and check measured input state before trusting the object-class label alone given the real contradictory M12/M3 evidence.

Controls and starting ranges

named shapes: mild/medium/strong S curves; highlight rolloff; shadow lift; globular_balanced; globular_core_rolloff

Mask behavior: Supported natively; current default Experiment Mode dispatch does not inject the generic adapted mask for pi_curves due to the key-name mismatch.

Failure modes

Inversion/wrong-knee behavior from a preset calibrated on a different input histogram.

Recovery

Revert; use monotonic measured custom points instead of a fixed object-class preset.

Seti Astro Suite Pro curves LUT curves_preset LUT machinery; project pin 1.18.0

Remember NOVA's shared FITS loader performs a global min/max normalization to [0,1] before this LUT is built -- that preprocessing is part of the effective treatment.
Treat channel="L" as a distinct, lower-precision strategy (8-bit OpenCV Lab quantization), not a high-bit-depth luminance-curve equivalent.

Controls and starting ranges

shape: linear / s_mild / s_med / s_strong / lift_shadows / crush_shadows / fade_blacks / rolloff_highlights
channel: all (linked) / R / G / B / L (8-bit Lab quantized)

Mask behavior: Supported through the generic nonlinear adaptation mask when the function name matches the mask-map key.

Failure modes

Treating channel="L" results as equivalent to a high-precision floating-point luminance curve.

Recovery

Use linked ("all") or a real luminance-preserving strategy where 8-bit quantization loss matters.

Siril curves/histogram tools native curves/histogram transformation; exact version not independently confirmed in this repo

Treat as the same conceptual family -- a pointwise brightness remapping -- with different interpolation/implementation than NOVA's SASpro or PixInsight paths.
Apply the same before/after measurement discipline (percentiles, clipping fraction, protected-ROI checks) used for NOVA's own candidates.

Controls and starting ranges

curve editor: manual control-point placement, typically post-stretch

Mask behavior: Not independently confirmed in this repo.

Failure modes

Treating a Siril curves result as evidence about NOVA's SASpro/PixInsight candidates.

Recovery

Prefer the no-op control if the result damages the target.

Validation status

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

Sources

Phase-2 research packet P27, 2026-08-28 packet:P27:2026-08-28 — NOVA source-confirmed
NOVA curves implementation, ontology candidates, and adaptation gap nas_server/seti_astro.py:curves; nas_server/tool_params.py:compute_curves; nas_server/processing_ontology.json curves; nas_server/experiments.py _adapt_nonlinear_variants — NOVA source-confirmed
M12/M3 real runs: globular core-rolloff curve crushed halo tones or sky without core benefit critiques (M12 2026-07-15 workflow 1.22.1; M3 2026-07-15 workflow 1.23.0) — NOVA execution record
PixInsight staff: what is a linear versus non-linear image — Vendor-documented
NIST/SEMATECH Engineering Statistics Handbook: randomized block designs — Vendor-documented
Ontology 'none' control candidates fail dispatch instead of running a true copy-through (issue #669) nas_server/experiments.py:_run_variant; SeeStar-db#669 — NOVA source-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.