Nonlinear Star Removal and Starless Finishing

Optionally and temporarily separate stellar from non-stellar content on an already-nonlinear (post-stretch) image, so extended structure -- nebulosity, galaxy detail, dust -- can be finished without simultaneously pushing, recoloring, or clipping stars.

schema 1 · revision 1

Recognize it

Extended-structure finishing operations (local contrast, dark-structure enhancement, selective saturation) keep interacting with stars because they are still merged into the same nonlinear image.
A finishing pass that helps nebulosity or galaxy detail simultaneously enlarges, recolors, or blows out stars, forcing an unwanted compromise.

Intended output

A nonlinear starless working image suitable for selected extended-object finishing. This is a model-conditioned reconstruction, not an observation of the sky with stars physically absent -- a clean-looking filled region is not recovered hidden astrophysical data, and compact non-stellar structure can be removed if the model classifies it as star-like.

Required input state

Nonlinear, already-stretched data with stable background modeling and color state, a completed stretch whose stellar profiles are not pathologically distorted, no gross clipping that makes star cores/halos impossible to model cleanly, and known scale/resampling state. RC-Astro documents that removal effectiveness can degrade with severe optical aberrations, excessive oversampling, and certain aggressive/alternative stretches such as GHS or arcsinh.

What NOVA does

NOVA's DarkStar path (`seti_astro.remove_stars_inprocess`) is a headless in-process SASpro wrapper accepting `mode` (default `unscreen`), `use_gpu` (default true), and `chunk_size` (default 512), writing a starless FITS. NOVA's StarXTerminator path (`seti_astro.star_removal_starxt`) imports `run_rcastro()` and dispatches operation `sxt` -- the same stand-alone RC-Astro CLI/RunPod GPU architecture used elsewhere in the pipeline, not a PixInsight-host process, despite the ontology candidate id `pi_starxt`; NOVA restores celestial WCS on the output afterward since RC-Astro's dispatcher drops it. The `none` candidate copies the input unchanged as a true control. None of these three receives NOVA's generic nonlinear adaptation/mask layer, since `star_removal` is absent from `_NL_STEPS`.

Use when

Extended target structure is the finishing priority and stars are currently constraining how far that finishing can go.
The parent stretch gives recognizable, reasonably clean stellar profiles -- not one of RC-Astro's documented degrading conditions (severe aberration, oversampling, GHS/arcsinh-shaped extreme stretches).
The target does not contain many compact features easily confused with stars, or a protective mask is available in the chosen workflow.
A true `none` control is retained in the comparison, not dropped for expedience.

Skip when

Stars are the subject -- clusters, star-rich fields, or any target where removing stars would remove the point of the image.
The target contains compact nonstellar structure at meaningful risk of misclassification (bright compact galaxy nuclei, planetary-nebula cores, cometary features, filament intersections) without an available protective mask.
The parent stretch or optical state falls into RC-Astro's documented degrading conditions for removal effectiveness.
Starless processing offers no specific downstream advantage for this target -- it is not a default 'better workflow' step.

Limits

Star removal is segmentation and inpainting, not physical signal recovery. This is the central scientific-language boundary for this article: nothing downstream should describe a starless result as having revealed data that was actually recorded underneath a star.
NOVA's four current candidates are `darkstar_unscreen` and `darkstar_additive` (SASpro DarkStar, same engine, different extraction mode -- strategy probes within one engine family, not independent algorithms), `pi_starxt` (RC-Astro StarXTerminator dispatched through NOVA's stand-alone RC-Astro CLI/RunPod path, restoring celestial WCS afterward since RC-Astro drops it), and `none`, a true unchanged control. The ontology's `pi_starxt` naming and its 'best-in-class' description are both stale/unsupported: current execution is the stand-alone RC-Astro CLI, not a PixInsight host process, and 'best-in-class' is vendor marketing language, not a repository-established comparative evidence level -- neither should be repeated here as validated fact.
`star_removal` is not in Experiment Mode's `_NL_STEPS`, so it receives none of the generic nonlinear luminance-mask injection or parameter adaptation that several neighboring nonlinear families do. That removes one common cross-candidate confounder, but real treatment differences remain: different AI models/training domains, DarkStar's unscreen-vs-additive semantics, RC-Astro-vs-SASpro preprocessing, tiling/chunking behavior, GPU/CPU numerical differences, and WCS restoration applying only to the StarXTerminator path. The absence of generic adaptation is not the same thing as full engine equivalence.
RC-Astro explicitly documents that StarXTerminator can remove non-stellar compact features -- comet cores, compact galaxy cores, and filamentary nebular structure are named vendor examples -- and that masks are sometimes required to protect them. A starless result therefore cannot be validated by a lower star count alone.
The `none` control is not a weak candidate. On star-rich targets, clusters, fields with difficult optical aberrations, or targets containing compact non-stellar structure, removal can do more harm than the starless finishing it was meant to enable -- `none` winning is a scientifically legitimate outcome, not a failed experiment.
This experiment evaluates the starless output alone. It does not validate a complete matching stars-only layer or a later recombination -- a good-looking starless result is not proof that a compatible stars-only layer and clean recombination exist.
This is a distinct operation from Linear Star Split (a separate article), which splits an already-linear parent before stretching using different extraction math for a linear image; the two should never be treated as the same strategy at different pipeline positions.

Scientific and aesthetic notes

The value of starless processing is workflow control and perceptual separation, not recovery of occluded information -- this framing should govern how any starless result is described, not just how it is judged.
A fair DarkStar-vs-StarXTerminator comparison needs the exact same nonlinear parent, the same crop/WCS/scale/resampling state, the same upstream stretch and star-correction history, recorded engine/model/tool version, a true unchanged `none` control, target-preservation regions of interest selected before results are known where practical, and artifact inspection at full/native resolution rather than preview thumbnails.
`darkstar_unscreen` vs. `pi_starxt` is not guaranteed to isolate 'DarkStar model vs. StarXTerminator model' in isolation -- their preprocessing and internal strategies differ, so treat any comparison as a complete-method comparison unless internal settings are demonstrated equivalent.
A winner measured on one stretch-engine-shaped parent (for example a GHS stretch) should not be generalized to a differently-shaped parent (for example Statistical Stretch or STF) without repeated evidence -- RC-Astro itself documents stretch shape as a factor in removal effectiveness.

Evidence and measurements

Measurements	Acceptance criteria
Residual stellar energy/profile test at matched star locations before/after, measuring residual compact flux/rings rather than only counting detected stars. Target-preservation regions of interest on known compact non-stellar structures, checked for local flux/structure loss. Difference image (input minus starless output), interpreted with nonlinear compositing math in mind, to reveal removed target structure, halos, seams, or residual star cores. Round-trip reconstruction error when a matching stars layer exists, recombined with the correct additive/screen rule and compared to the parent. Tile/seam inspection, especially on large or high-resolution/drizzled images. Background statistics in genuinely unchanged regions, to detect model-induced texture or noise drift. Color residual inspection for cyan/magenta/red rings or color holes around removed stars.	Residual stellar energy at former star positions is measurably reduced without new dark holes, bright rings, or duplicated texture. Known compact non-stellar structures retain their measured local flux and morphology. No systematic tile seams or background texture/noise drift in otherwise-unchanged regions. The candidate actually executed (engine, mode, and whether `none` won) is recorded -- a lower star count, smoother background, or more dramatic-looking target is never accepted alone as validation. `none` winning is recorded as a legitimate outcome, not reported as a failed or incomplete experiment.

Measurements

Acceptance criteria

Residual stellar energy/profile test at matched star locations before/after, measuring residual compact flux/rings rather than only counting detected stars.
Target-preservation regions of interest on known compact non-stellar structures, checked for local flux/structure loss.
Difference image (input minus starless output), interpreted with nonlinear compositing math in mind, to reveal removed target structure, halos, seams, or residual star cores.
Round-trip reconstruction error when a matching stars layer exists, recombined with the correct additive/screen rule and compared to the parent.
Tile/seam inspection, especially on large or high-resolution/drizzled images.
Background statistics in genuinely unchanged regions, to detect model-induced texture or noise drift.
Color residual inspection for cyan/magenta/red rings or color holes around removed stars.

Residual stellar energy at former star positions is measurably reduced without new dark holes, bright rings, or duplicated texture.
Known compact non-stellar structures retain their measured local flux and morphology.
No systematic tile seams or background texture/noise drift in otherwise-unchanged regions.
The candidate actually executed (engine, mode, and whether `none` won) is recorded -- a lower star count, smoother background, or more dramatic-looking target is never accepted alone as validation.
`none` winning is recorded as a legitimate outcome, not reported as a failed or incomplete experiment.

Cross-tool matrix

Tool	Equivalence	Expected result	Evidence
nova	Exact replay	A nonlinear starless working image with residual stellar energy at former star positions minimized without measurable loss of compact non-stellar target structure, or a `none` result when removal would cost more than it helps for this target.	nova-starless-finishing-source
pixinsight	Same engine, adapted host	A starless image with minimized residual star energy and no material loss of known compact non-stellar structure.	rc-astro-sxt-usage-notes-p31, rc-astro-sxt-faq-failure
saspro	Functional alternative	A starless working image via DarkStar's chosen extraction mode.	saspro-darkstar-source-p31, nova-starless-finishing-source
siril	Functional alternative	A starless image and a stars-only companion via a different star-removal model family than either of NOVA's current engines.	nova-starless-finishing-source

NOVA Python pipeline ontology current; RC-Astro CLI/RunPod GPU dispatch for pi_starxt, SASpro in-process for DarkStar

Confirm the parent is genuinely nonlinear/post-stretch before running this step -- it is not interchangeable with Linear Star Split, which requires a linear parent and uses different extraction math.
Do not read the `pi_starxt` candidate id as implying a PixInsight-host execution; verify it dispatched through the stand-alone RC-Astro CLI/RunPod path if provenance matters.
Retain the `none` control in any comparison rather than assuming removal is always preferable.

Controls and starting ranges

darkstar_unscreen / darkstar_additive: mode unscreen or additive; GPU on by default
pi_starxt: stand-alone RC-Astro StarXTerminator via run_rcastro('sxt'); WCS restored on output
none: unchanged copy of the input; true control

Mask behavior: Not currently wired into NOVA's headless dispatch of this step for either engine.

Failure modes

Presenting the ontology's 'best-in-class' StarXTerminator description as an established comparative result rather than vendor marketing language.
Judging a candidate by star count or a smoother-looking background alone.
Confusing this step's output with Linear Star Split's starless output.

Recovery

Fall back to `none` (unchanged parent) if all removal candidates damage the target or introduce unacceptable artifacts.

RC-Astro StarXTerminator (stand-alone, headless -- no PixInsight host process) StarXTerminator, stand-alone RC-Astro CLI path; exact model revision not independently confirmed in this repo

Apply to an already-nonlinear/stretched image, consistent with RC-Astro's own documented workflow position for this use case.
Use a protective mask for known compact nonstellar structures where the workflow supports it.
If removal appears to fail or leave residuals, check for RC-Astro's documented degrading conditions (severe aberration, oversampling, unusual stretch shape) before assuming a model defect.

Controls and starting ranges

documented degrading conditions: severe optical aberration; excessive oversampling; certain aggressive/alternative stretches (GHS, arcsinh)

Mask behavior: RC-Astro documents mask support for protecting non-stellar compact structure; not currently wired into NOVA's headless dispatch of this step.

Failure modes

Assuming a starless result validates itself without checking target-preservation regions of interest.

Recovery

Correct upstream conditions where possible, or prefer the `none` control for this target.

Seti Astro Suite Pro DarkStar DarkStar/Cosmic Clarity; exact NOVA-deployed model version not independently confirmed in this repo

Choose unscreen or additive mode deliberately -- these are different extraction/recombination assumptions, not merely a strength setting, and should be documented as separate strategies.
Compare against a true `none` control, not only against the other removal engine.

Controls and starting ranges

mode: unscreen or additive
use_gpu: true by default
chunk_size: 512 (implementation/performance control)

Mask behavior: Not currently exposed by NOVA's wrapper for this step.

Failure modes

Treating unscreen and additive as interchangeable strength settings of the same strategy.

Recovery

Prefer the `none` control if both DarkStar modes damage the target.

Siril native StarNet nonlinear star removal native StarNet integration; exact version not independently confirmed in this repo

Siril's built-in `starnet` command is a closer conceptual match here than for the linear split, since it is typically used on already-stretched data -- but it is still a different model family (StarNet++) from both DarkStar and StarXTerminator, so treat it as a functional alternative, not a matched engine.
Apply the same target-preservation and residual checks used for NOVA's own candidates before accepting a StarNet result -- no engine is exempt from the compact-structure-loss risk.
Record the exact Siril version and command options; this integration's removal quality has not been independently verified against NOVA's pipeline in this repo.

Controls and starting ranges

availability: native `starnet` command, typically applied post-stretch

Mask behavior: Not independently confirmed in this repo.

Failure modes

Treating a Siril StarNet result as evidence about DarkStar or StarXTerminator quality.

Recovery

Prefer the `none` control if StarNet's output damages the target.

Validation status

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

Sources

Phase-2 research packet P31, 2026-08-28 packet:P31:2026-08-28 — NOVA source-confirmed
NOVA nonlinear star removal implementation and Experiment Mode candidates nas_server/seti_astro.py:remove_stars_inprocess,star_removal_starxt; nas_server/processing_ontology.json star_removal; nas_server/experiments.py _NL_STEPS — NOVA source-confirmed
RC-Astro StarXTerminator Usage Notes — Vendor-documented
RC-Astro: Why does StarXTerminator fail to remove stars? — Vendor-documented
Seti Astro Suite Pro DarkStar / 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.