Stacking, Integration, Weighting, and Rejection

Combine a declared set of registered exposures so persistent astronomical signal accumulates while normalization, weighting, and pixel-level outlier rejection limit noise and transient defects.

schema 1 · revision 1

Recognize it

A single exposure is noisy, but the target repeats at the same registered coordinates across frames.
Satellite trails, cosmic rays, or residual defects affect isolated pixel samples rather than every exposure.
Different integration strategies produce different noise, detail, artifact, or coverage tradeoffs from the same frame set.

Intended output

One still-linear integrated science image with its exact input manifest, frame count, total exposure, output grid, normalization, weighting, rejection, and executed engine recorded.

Required input state

A frozen, provenance-bearing light-frame manifest after frame-level subframe inspection and culling.
Frames registered to a declared common geometry, with calibration, CFA/debayer, filter, exposure, mount mode, panel, and session compatibility known.
A declared output grid, framing, drizzle state, normalization goal, weighting policy, rejection policy, and common measurement support.

What NOVA does

NOVA currently exposes four meaningful stack paths. Siril is the reference strategy: Siril registration followed by a 3/3 Winsorized-rejection mean, additive-scale normalization, background-noise weighting, and output normalization. ImageMM uses Siril-registered frames for a multi-frame deconvolution stack, currently 20 iterations, kappa 2.0, PerChannel. The historically named pixinsight_wbpp path is actually a hybrid: Siril registration plus PixInsight ImageIntegration with 3/3 Winsorized rejection, additive scaling, and requested SNR weighting. pixinsight_register changes registration, grid, and integration together. Requested behavior can fall back or change, so comparisons must record the executed treatment rather than only the requested engine name.

Use when

Two or more compatible, registered exposures contain repeated target signal and partly independent noise.
A declared rejection and weighting policy is supportable for the available frame count and data class.
Alternative engines are being compared from a frozen manifest with state differences controlled or explicitly labeled.

Skip when

There is only one usable exposure; no-op is not a comparable stacking control, so use a declared reference integration strategy instead.
Calibration, membership, registration, CFA/channel state, or output geometry is unresolved.
The sample count is too small for the proposed rejection rule to estimate outliers reliably.
The intended comparison cannot separate integration behavior from changed registration, drizzle, framing, or measurement support.

Limits

Frame-level culling and pixel-level rejection are different decisions: subframe inspection decides whether an entire exposure enters the stack; rejection decides whether an individual aligned pixel sample is an outlier during integration.
Rejection needs enough comparable samples to estimate a distribution; aggressive thresholds can remove real faint or thin structure as well as transients.
The ideal random-noise improvement is approximately proportional to the square root of comparable exposure count, but gradients, seeing, correlated noise, resampling, and unequal frames prevent that from being a guarantee.
An engine comparison that changes frame selection, registration grid, drizzle, framing, normalization, or support is a strategy/bundle comparison, not proof that one pixel-combination algorithm is intrinsically better.
NOVA's stack SNR is a structure-to-background-noise proxy, not conventional astronomical photometric SNR; efficiency derived from it is not a universal truth score.
A readable output proves artifact production, not correctness, preservation, or a class-level engine preference.

Scientific and aesthetic notes

Subframe Inspection and Culling is the frame-level decision page; this page begins with its retained manifest and does not duplicate its scoring guidance.
Measuring an Astrophoto defines why matched support and a vector of measurements matter; no single sharpness, SNR-proxy, or efficiency value should rank all stack strategies.
ImageMM couples integration with restoration, so smaller FWHM or more high-frequency structure can reflect ringing, noise amplification, or sharpened artifacts rather than more truthful detail.
Drizzle, max-framing mosaics, and canonical grids change sampling or support; compare FWHM in common angular units and keep coverage performance separate from common-support image quality.
A class-level preference requires repeated, version-applicable evidence across independent targets and data conditions, not one attractive result or a historical win fraction.

Evidence and measurements

Measurements	Acceptance criteria
Exact included frame identities, count, total exposure, and spatial coverage. Robust background-noise proxy on common sky support, with sampling and resampling state recorded. FWHM and eccentricity from a stable star population in common angular or grid units. Residual gradients, seams, clipping/pile-up diagnostics, and rejected-sample artifacts. Fine-structure preservation plus ringing, false-detail, and noise-amplification inspection. Runtime, memory, storage, I/O, and failures reported separately from image quality.	The artifact is readable, linear, on the declared grid, and accounts for the expected manifest and exposure. Frame-level culling is explicitly distinguished from pixel-level rejection throughout the evidence and explanation. No catastrophic registration, coverage, clipping, rejection-hole, trail, ringing, or preservation failure is accepted. Comparisons use common support and disclose every unavoidable registration, drizzle, framing, normalization, weighting, and rejection difference. SNR-proxy and efficiency values remain qualified; no one-dimensional score or single dataset establishes a universal engine preference.

Measurements

Acceptance criteria

Exact included frame identities, count, total exposure, and spatial coverage.
Robust background-noise proxy on common sky support, with sampling and resampling state recorded.
FWHM and eccentricity from a stable star population in common angular or grid units.
Residual gradients, seams, clipping/pile-up diagnostics, and rejected-sample artifacts.
Fine-structure preservation plus ringing, false-detail, and noise-amplification inspection.
Runtime, memory, storage, I/O, and failures reported separately from image quality.

The artifact is readable, linear, on the declared grid, and accounts for the expected manifest and exposure.
Frame-level culling is explicitly distinguished from pixel-level rejection throughout the evidence and explanation.
No catastrophic registration, coverage, clipping, rejection-hole, trail, ringing, or preservation failure is accepted.
Comparisons use common support and disclose every unavoidable registration, drizzle, framing, normalization, weighting, and rejection difference.
SNR-proxy and efficiency values remain qualified; no one-dimensional score or single dataset establishes a universal engine preference.

Cross-tool matrix

Tool	Equivalence	Expected result	Evidence
nova	Exact replay	A reproducible linear integration with manifest, executed treatment, and quality/resource evidence retained.	nova-stacking-source
pixinsight	Functional alternative	A weighted/rejected linear integration whose registration and parameter differences are explicit.	nova-stacking-source
siril	Same engine, adapted host	A conventional robust weighted integration suitable as NOVA's reference strategy.	siril-stacking-1.4.4, nova-stacking-source
saspro	Same engine, adapted host	A restored multi-frame integration assessed for both detail and preservation artifacts.	saspro-imagemm-vendor, nova-stacking-source

NOVA stack dispatch current source snapshot described by P09

Freeze the eligible frame manifest after frame-level culling and registration decisions.
Select a reference or alternative strategy and record requested plus executed engine, versions, framing, drizzle, grid, normalization, weighting, rejection, and fallback behavior.
Inspect common-support measurements, coverage, rejected samples or maps when available, and artifacts before accepting the integrated result.

Controls and starting ranges

Siril reference: Winsorized 3/3; additive-scale normalization; background-noise weighting
ImageMM current runtime: 20 iterations; kappa 2.0; PerChannel
hybrid PI integration: Winsorized 3/3; additive scaling; requested SNR weighting; avgdev or hero-mode IKSS scale

Mask behavior: Use one declared common celestial/support region for comparisons; coverage maps and low-coverage edges are assessed separately.

Failure modes

Frame-level exclusions are mistaken for pixel rejection.
A fallback silently changes the requested strategy.
One proxy or sharpened appearance is treated as an overall engine verdict.

Recovery

Return to the frozen manifest and a conventional Siril reference integration.
Match or explicitly disclose grid, support, drizzle, and framing differences.
Inspect rejection artifacts and preservation evidence before adjusting thresholds.

PixInsight ImageIntegration source-recorded 1.9.3 behavior; installed version unverified

Use the same registered frame manifest and output geometry when testing integration alone.
Match normalization, rejection, and weighting goals as closely as the tool permits and record differences.
Inspect rejection maps and common-support output; call a PI-native registration run a strategy comparison, not integration-only evidence.

Controls and starting ranges

NOVA hybrid reference: Winsorized 3/3, additive scaling, SNR weight request

Mask behavior: Measure a common valid region; retain rejection maps separately from the integrated image.

Failure modes

Calling NOVA's hybrid path full WBPP.
Changing registration or grid while claiming an integration-only comparison.

Recovery

Retain Siril registration for the hybrid comparison.
Relabel PI-native registration plus integration as a bundle comparison.

Siril sequence stacking 1.4.4 documentation; NOVA source targets 1.4.x

Stack the frozen registered sequence with the declared combination, normalization, weighting, rejection, and framing policy.
Record frame count and generate rejection maps when inspection is needed.
Compare on common support and do not treat the current 3/3 thresholds as universal optima.

Controls and starting ranges

NOVA reference: rej 3 3, implicit Winsorized, addscale, noise weighting, output normalization

Mask behavior: Use common valid support for quality measurements; -maximize union coverage is a separate declared treatment.

Failure modes

Too few samples make clipping unstable.
Max-framing blank or low-coverage regions dominate whole-image statistics.

Recovery

Reduce or disable rejection for a small sample and inspect the result.
Use a common-support mask and assess union coverage separately.

Seti Astro Suite Pro ImageMM vendor capability documented; installed version unverified

Use the same registered manifest and record the actual iteration, kappa, and color-mode settings.
Treat ImageMM as integration plus multi-frame restoration rather than another robust mean.
Inspect stars, faint structure, and high-contrast edges for ringing, false detail, and amplified noise.

Controls and starting ranges

NOVA current runtime: 20 iterations; kappa 2.0; PerChannel

Mask behavior: Use common support and inspect artifact-prone star edges and faint structures at native scale.

Failure modes

Sharpness-oriented metrics reward the method by construction.
The ontology's unexecutable 40-iteration candidate is reported as current behavior.

Recovery

Compare against the conventional reference on common support.
Return to the source-confirmed 20-iteration runtime treatment and record executed parameters.

Validation status

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

Sources

Phase-2 research packet P09, 2026-08-27 Handbook P09 — Stacking, Integration Engines, and Rejection Strategy — Reasoned translation
NOVA stack dispatch, configuration, and assessment source nas_server/stacker.py:stack_target; nas_server/stack_config.py; nas_server/seestar_stack_only.ssf; nas_server/stack_assessor.py — NOVA source-confirmed
Siril 1.4.4 stacking command documentation — Vendor-documented
Seti Astro Suite Pro ImageMM stacking description — Vendor-documented

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.