Post-Stretch Denoise

Suppress visually objectionable residual noise -- stochastic grain, chroma speckle, correlated resampling texture -- after the main stretch, while preserving real low-contrast structure and compact features. A nonlinear stretch changes the apparent amplitude and spatial distribution of noise that linear-stage denoise (a separate article) already addressed once; this is an optional second pass on a different signal/noise regime, not a repeat of the same step.

schema 1 · revision 1

Recognize it

Dark background variation, faint-detail grain, or chroma speckle became visually prominent after the main stretch, even though color calibration and linear-stage denoise are correct.

Intended output

Reduced objectionable stochastic grain/chroma speckle in matched background and faint-signal regions, with low-contrast filaments, dust, galaxy outskirts, small-scale target texture, and star profiles/cores preserved -- no plastic, waxy, patchy, tiled, or smeared texture, and no unjustified change to background level, gradients, hue, or dynamic-range mapping.

Required input state

A nonlinear, already-stretched image, past the main linear restoration chain (deconvolution, linear denoise, star correction as applicable), with residual noise actually assessed as objectionable rather than denoised reflexively. NOVA workflows place this step after stretch and other nonlinear cleanup/contrast operations, so the real parent can be a post-background-neutralize or post-CLAHE product depending on the active workflow path -- CLAHE/local contrast can amplify noise, making a subsequent pass here qualitatively different from denoise applied immediately after stretch alone.

What NOVA does

NOVA's production default is `denoise_nxt()`: a second, conservative NoiseXTerminator pass at `nxt_denoise=0.65`, `nxt_iterations=2`, dispatched through the remote RC-Astro GPU when enabled and the local RC-Astro CLI otherwise -- never silently substituting Cosmic Clarity. This is distinct from the pre-stretch linear NXT pass, which remains unchanged. Cosmic Clarity (`cc_denoise_inprocess()`, separate luminance/color strengths) and Prism Mini (`syqon_prism_denoise()`, a pure in-process NumPy path with explicit tile/overlap/pad controls) remain available as explicit Experiment Mode candidates, though the Cosmic Clarity candidates currently fail to dispatch (#679). `scripts/denoise_bakeoff.py` provides a real four-arm comparison scaffold (unchanged baseline, Cosmic Clarity, Prism Mini, NXT) from a shared post-stretch parent for evaluation outside the standard pipeline path.

Use when

The stretched image shows objectionable stochastic grain or chroma speckle after the main stretch, and denoise strength can be checked against real preservation of faint structure rather than applied automatically.

Skip when

The image is already adequately denoised from the linear stage and shows no new objectionable noise after stretch -- a second pass risks compounding detail loss for no real gain.
The workflow's own object-type routing omits this step for the target class (the galaxy workflow currently omits the nonlinear denoise step entirely, per workflow 1.25.0's own changelog note).
The parent state includes strong prior local-contrast enhancement (CLAHE, DSE) that has already amplified noise texture in a way this step cannot cleanly separate from real structure.

Limits

Denoising does not recover information that was never recorded. A smoother result is not automatically a more truthful one -- learned denoisers can suppress real faint texture as noise, and strong settings can create plasticky/smeared regions or edge discontinuities. RC Astro's own NXT documentation makes this explicit: 0 means no reduction, 1 means attempting full reduction, and 100% removal can look unnaturally smooth.
NoiseXTerminator is NOVA's current production default for this step, not one of three roughly-equal engines. Workflow 1.25.0 (2026-08-29) made `denoise_nxt` the ontology's `seti_astro_fn`, replacing the prior default `cc_denoise_inprocess`, following a real M31 broadband benchmark where a full local Cosmic Clarity run took approximately 27.4 minutes and the objective gate rejected the resulting candidate anyway. Cosmic Clarity and Prism Mini remain real Experiment Mode alternatives, not the production default -- describing them as three symmetric production choices overstates their current standing.
The ontology's `cc_denoise_gentle` and `cc_denoise_moderate` candidates declare `fn: "cc_denoise"`, which does not match any real `seti_astro` function (the real function is `cc_denoise_inprocess`). Both candidates currently fail every time they are selected, returning `ok: False, error: "seti_astro.cc_denoise not found"` instead of running Cosmic Clarity (issue #679). Until that ontology fix lands, Cosmic Clarity is not a working Experiment Mode alternative in practice, regardless of what the candidate list appears to offer.
Cross-engine numeric strengths are not commensurate: Cosmic Clarity's luma/color values, Prism Mini's `strength`, and NXT's denoise/iterations represent different controls on different models. Comparing one preset from each is a complete-strategy comparison, not an isolated test of engine superiority -- do not translate a strength value from one engine onto another.
Prism Mini's tile/overlap/pad geometry (fixed at 512/64/64 in current ontology candidates) is part of treatment identity, not an implementation detail -- fixed pixel-space tile geometry maps to different angular scales at different drizzle factors, and seams at tile boundaries are a real failure mode worth an explicit acceptance check, especially on large or mosaic frames.
A prior Handbook/source audit recorded an NXT treatment-identity risk: the ontology includes an `nxt_detail` field, but the candidate's own description now states outright that "the RC-Astro CLI does not currently expose an equivalent control" -- the field is recorded for preset parity, not necessarily consumed by the runtime. Effective executed parameters, not ontology labels, are the real treatment identity.
`scripts/denoise_bakeoff.py` is a real, valuable execution/comparison scaffold -- it runs a true unchanged baseline, Cosmic Clarity, Prism Mini, and NXT from the same post-stretch starless parent and records `snr`, background RMS, median, and a `grain = background_rms / median` ratio. That proves the comparison workflow is designed and executable; it does not by itself establish a validated engine ranking, and its metrics inherit the same caveats as the rest of this family (see measurement traps below).

Scientific and aesthetic notes

A rigorous experiment should start from the same parent FITS, hold stretch/curves/CLAHE/HDR/color/star-state/cropping fixed, include a true unchanged baseline, record engine/model/tool version and actual executed parameters (not just candidate ID), record masks/tiling/normalization behavior and whether the image scale is native/drizzled/mosaic, compare multiple target/background structures rather than whole-frame statistics alone, and inspect full-resolution outputs and difference images rather than JPEG previews. If claiming a cross-engine comparison, tune engines fairly rather than comparing arbitrary preset numbers as if the scales were equivalent.
A within-engine strength result (e.g. NXT gentle vs. standard vs. strong) and a cross-engine method result (NXT vs. Cosmic Clarity vs. Prism Mini) are two different levels of interpretation. Pooling both into one historical "noise_reduction winner percentage" erases treatment-identity differences that matter.
Post-stretch denoise is highly parent-state dependent: the same candidate run after stronger CLAHE/curves or a different stretch has a different apparent need and different artifact behavior. Historical winner rates across unlike parent states are weak evidence unless blocked by stretch, local-contrast history, target class, image scale, and prior denoise state.
"Plasticity" -- waxy smooth regions, suppressed faint granular structure, local texture inconsistency -- is not captured by one scalar. Entropy or high-frequency power can detect that something changed but cannot identify whether the removed high-frequency content was noise or real signal. Human/perceptual inspection remains necessary, paired with matched structural ROIs and difference images, not treated as unstructured taste alone.

Evidence and measurements

Measurements	Acceptance criteria
Robust sky/background RMS or MAD reduction, and chroma-noise reduction in neutral/background regions, on matched parent/output states and appropriate ROIs. Local variance reduction in designated empty-sky ROIs, and brightness-stratified star-profile change if stars are present. Structural residual/difference images in faint target ROIs, and preservation of known filament/dust boundaries across matched coordinates. Tile-boundary discontinuity checks for tiled engines (Prism Mini). Runtime/resource cost, since it materially informed the real production engine decision (the M31 27.4-minute Cosmic Clarity benchmark).	Objectionable residual noise is reduced in matched background/faint-signal regions while low-contrast filaments, dust, galaxy outskirts, target texture, and star profiles are preserved, checked against a real unchanged baseline. No waxy/plastic texture, tile seams, or chroma smearing introduced. Effective engine, model/tool version, and actual executed parameters are recorded -- not just candidate ID, given the nxt_detail forwarding caveat and the currently-broken Cosmic Clarity candidates. Whole-frame NOVA SNR-like values, background_rms/median grain ratio, entropy, high-frequency power, generic sharpness, SSIM against the noisy parent, and AI/perceptual preference are never presented as standalone proof -- lower RMS can reward deleting faint signal, and a higher SNR-like proxy can reward over-smoothing. A cross-engine result is never presented as isolating engine superiority unless effective strengths were fairly matched or each engine received its own fair tuning opportunity.

Measurements

Acceptance criteria

Robust sky/background RMS or MAD reduction, and chroma-noise reduction in neutral/background regions, on matched parent/output states and appropriate ROIs.
Local variance reduction in designated empty-sky ROIs, and brightness-stratified star-profile change if stars are present.
Structural residual/difference images in faint target ROIs, and preservation of known filament/dust boundaries across matched coordinates.
Tile-boundary discontinuity checks for tiled engines (Prism Mini).
Runtime/resource cost, since it materially informed the real production engine decision (the M31 27.4-minute Cosmic Clarity benchmark).

Objectionable residual noise is reduced in matched background/faint-signal regions while low-contrast filaments, dust, galaxy outskirts, target texture, and star profiles are preserved, checked against a real unchanged baseline.
No waxy/plastic texture, tile seams, or chroma smearing introduced.
Effective engine, model/tool version, and actual executed parameters are recorded -- not just candidate ID, given the nxt_detail forwarding caveat and the currently-broken Cosmic Clarity candidates.
Whole-frame NOVA SNR-like values, background_rms/median grain ratio, entropy, high-frequency power, generic sharpness, SSIM against the noisy parent, and AI/perceptual preference are never presented as standalone proof -- lower RMS can reward deleting faint signal, and a higher SNR-like proxy can reward over-smoothing.
A cross-engine result is never presented as isolating engine superiority unless effective strengths were fairly matched or each engine received its own fair tuning opportunity.

Cross-tool matrix

Tool	Equivalence	Expected result	Evidence
nova	Exact replay	Reduced objectionable grain/chroma speckle in matched background and faint-signal regions with low-contrast structure, star profiles, and background level preserved.	nova-post-stretch-denoise-source, nova-nxt-default-decision, nova-cc-denoise-dispatch-bug
pixinsight	Same engine, adapted host	Suppressed residual noise via NXT's learned model at the configured denoise/iteration settings.	rc-astro-nxt-manual, nova-post-stretch-denoise-source
saspro	Functional alternative	Learned-model noise suppression via Cosmic Clarity or Prism Mini's in-process NumPy path.	setiastro-cosmic-clarity, nova-cc-denoise-dispatch-bug, nova-post-stretch-denoise-source
siril	Functional alternative	Reduced residual noise via Siril's own denoise tooling.	nova-post-stretch-denoise-source

NOVA Python pipeline denoise_nxt() production default (workflow 1.25.0); cc_denoise_inprocess()/syqon_prism_denoise() Experiment Mode alternatives; denoise_bakeoff.py comparison scaffold

Treat NXT as the deliberate production default, not one of three symmetric choices -- the 2026-08-29 switch away from Cosmic Clarity was a real, evidenced decision (M31 benchmark: ~27.4min runtime, objective-gate rejection), not an arbitrary preference.
Do not select cc_denoise_gentle/moderate expecting a working Cosmic Clarity comparison until #679 is fixed -- both currently fail dispatch.
Record actual executed parameters, not just candidate ID -- the ontology's nxt_detail field may not be consumed by the runtime path even when declared on a candidate.
For Prism Mini, record tile/overlap/pad alongside strength -- geometry is part of treatment identity, especially across native vs. drizzled image scale.

Controls and starting ranges

nxt_gentle / nxt_standard / nxt_strong: NXT denoise 0.45 / 0.65 / 0.80 -- production default runs at 0.65, 2 iterations
cc_denoise_gentle / cc_denoise_moderate: Cosmic Clarity luma/color 0.50/0.35 and 0.70/0.50 -- currently non-dispatching, see #679
prism_mini_gentle / prism_mini_standard: Prism Mini strength 0.65 / 0.85, tile 512 / overlap 64 / pad 64
none: true control

Mask behavior: Not currently exposed by NOVA's wrapper for any of the three engines in this family.

Failure modes

Waxy/plastic faint regions or loss of filaments/dust from strength set too high or a second denoise pass compounding detail loss.
Chroma smearing from color-denoise strength set too high.
Tile seams or local discontinuities on large/drizzled/mosaic frames from Prism Mini's fixed pixel-space tile geometry.
Treating a Cosmic Clarity candidate selection as a real comparison arm when it is currently failing to dispatch.

Recovery

Revert to the parent/unchanged control; reduce strength; separate luminance and chroma treatment; increase tile overlap/padding or switch engine for seam artifacts.

RC-Astro NoiseXTerminator 2/AI3 RC-Astro NoiseXTerminator via headless CLI/RunPod GPU dispatch, not a PixInsight host process

Vendor semantics: `denoise` amount 0 = no reduction, 1 = attempt full reduction; 100% removal can look unnaturally smooth. `iterations` describes successive-approximation steps -- more can retain detail better in noisy regions, but high counts can create artifacts.
NOVA's simpler candidate family does not expose NXT's full documented parameter space (e.g. intensity/color and high-/low-frequency separation modes in the full tool) -- a NOVA result is evidence about NOVA's selected configuration, not NXT as an abstract universal method.

Controls and starting ranges

denoise: 0-1, NOVA production default 0.65
iterations: successive-approximation steps, NOVA production default 2

Mask behavior: Not exposed by NOVA's current wrapper.

Failure modes

Assuming NOVA exercises NXT's full documented parameter space when only denoise/iterations are exposed.

Recovery

Reduce denoise amount or iterations; compare against the unchanged baseline.

Seti Astro Suite Pro Cosmic Clarity / Prism Mini Cosmic Clarity (in-process, currently non-dispatching for this family's ontology candidates -- #679) + Prism Mini

Cosmic Clarity exposes separate luminance and color denoise strengths -- meaningful because chroma speckle can be objectionable even when luminance structure should be preserved, though strong chroma reduction can smear legitimate weak color variation.
The installed SASpro package/model version is part of treatment identity for an AI denoiser embedded through SASpro -- treat exact model identity as version-bound unless captured by run provenance.
Do not expect cc_denoise_gentle/moderate to actually run until #679's ontology fix lands.

Controls and starting ranges

denoise_luma / denoise_color: 0.50/0.35 gentle, 0.70/0.50 moderate in current ontology candidates
Prism Mini strength: 0.65 gentle, 0.85 standard
Prism Mini tile / overlap / pad: 512 / 64 / 64 -- fixed pixel-space geometry, a reproducibility factor across native/drizzle scale

Mask behavior: Not exposed by NOVA's current wrapper for either engine.

Failure modes

Comparing Cosmic Clarity/Prism Mini strength values directly against NXT's denoise/iterations as if numerically equivalent.

Recovery

Reduce strength; separate luma/color treatment for Cosmic Clarity; adjust tile/overlap for Prism Mini seam artifacts.

Siril native denoise tools native denoise tools; exact version not independently confirmed in this repo

Treat as the same conceptual goal -- residual noise suppression on a nonlinear image -- with different implementation than NOVA's NXT/Cosmic Clarity/Prism Mini paths.
Apply the same joint noise-reduction-plus-structure-preservation measurement discipline used for NOVA's own candidates.

Controls and starting ranges

native denoise tools: typically applied post-stretch

Mask behavior: Not independently confirmed in this repo.

Failure modes

Treating a Siril denoise result as evidence about NOVA's NXT/Cosmic Clarity/Prism Mini candidates.

Recovery

Prefer the unchanged baseline if the result damages faint structure.

Validation status

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

Sources

Phase-2 research packet P29, 2026-08-28 packet:P29:2026-08-28 — NOVA source-confirmed
NOVA post-stretch denoise implementation: denoise_nxt, cc_denoise_inprocess, syqon_prism_denoise, denoise_bakeoff.py nas_server/seti_astro.py:denoise_nxt,cc_denoise_inprocess,syqon_prism_denoise; nas_server/processing_ontology.json noise_reduction; scripts/denoise_bakeoff.py — NOVA source-confirmed
Workflow 1.25.0: NXT chosen as the standard post-stretch engine over Cosmic Clarity, following a real M31 benchmark critiques/WORKFLOW_CHANGELOG.md (workflow 1.25.0, 2026-08-29); commit 1c63d87 — NOVA execution record
cc_denoise_gentle/moderate candidates reference a nonexistent seti_astro.cc_denoise function (issue #679) nas_server/processing_ontology.json noise_reduction; SeeStar-db#679 — NOVA source-confirmed
RC Astro NoiseXTerminator 2/AI3 User Manual (PixInsight) — Vendor-documented
Seti Astro Cosmic Clarity — 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.