Narrowband Imaging

Purpose, audience, prerequisites, and expected result

Use this workflow to turn mono emission-line or narrowband OSC captures into traceable mapped-color images. It is for users who can classify calibration and light frames, inspect linear images, and distinguish a display stretch from a pixel-changing stretch. It gives concrete starting points and checkpoints, not a promise that one palette or frame count suits every target.

You need:

• mono Ha/OIII/SII or other clearly identified filter data, or a calibrated and demosaiced narrowband OSC master whose passbands you understand;
• calibration frames matched to camera mode, exposure/temperature policy, binning, and filter/optical train;
• consistent filter labels, geometry, session, gain/ISO, offset/readout mode, object, and CFA metadata where applicable;
• a writable output directory outside the source tree, with room for masters, registered lights, history, reports, and optional diagnostics; and
• for continuum or broadband-star branches, registered same-target broadband data on the same pixel grid.

Expected branch products are:

Branch	Required linear inputs	Branch product and preservation point
Mono HOO	Ha + OIII masters	Registered/matched channels and an HOO composition; retain both mono masters
Mono SHO/custom	Ha + OIII + SII masters	Registered/matched channels and SHO, Foraxx, or custom mapped-color variants; retain all mono masters
Narrowband OSC	Calibrated/demosaiced RGB master, optionally direct mono roles	A palette-composer result after explicit OSC1/OSC2 role review; retain the original RGB master
NBRGB continuum enhancement	Registered linear RGB plus one or more mono narrowband masters	An RGB result with selected channels enhanced; retain RGB and narrowband inputs
Star-separated finish	Stretched color image plus a verified star mask/star image or optional external extraction	Matched starless/stars products and a recombined master; retain the pre-separation image

Verification data and evidence mode

Item	Record
Project-owned sample set	The mono SHO data set of M16 captured on 2025-06-18: maintainer-supplied project-owned mono Ha/OIII/SII lights and per-filter flats. Permission covers validation and retention of sanitized settings, hashes, observations, and derived evidence.
Branches covered	Physical mono SHO path: eight lights plus eight flats per Ha/OIII/SII group, 48 selected inputs. Direct SHO/HOO/Foraxx/custom, matching, native star, export, recovery, and negative paths were executed in the historical partial run. OSC1 and NBRGB arithmetic used same-target derivatives; representative physical OSC and matched broadband-continuum branches remain blocked.
External tools/models	None. StarNet, DeepSNR, and other optional executables/models were not used.
Catalogs/network	None. Gaia enrichment, online solving, response packs, catalogs, and network access were disabled/not used.
Platform/application build	Historical partial execution: Windows MSVC x64 Release and Debug, Qt 6.10.3. The blocked-remediation numbered builds and complete-suite results are retained separately in the evidence record.
Evidence artifacts	Sanitized frame IDs/hashes, historical partial-run binary/source/output/report hashes, exact settings, stage results, warnings, bounded image observations, export/reopen, recovery evidence, exact focused commands, and blocked-remediation validation are in the guide's validation record. Private paths, reports, logs, manifests, and generated images remain only in isolated approved output trees.
Unverified claims excluded	Physical narrowband OSC acquisition/calibration/demosaic/sensor behavior and matched broadband-continuum production behavior remain unverified, so verification of this workflow is incomplete. No claim is made that eight frames or recorded values are production-optimal; that uncalibrated flats are recommended; or that external tools, other platforms, polished aesthetics, or printing were accepted.

The retained mode-E run proves only the recorded mono branch and bounded proxy artifacts. It does not satisfy the mandatory representative-data gate for the OSC and continuum branches. Apply the checkpoints to your data instead of copying the values blindly.

Inputs and outputs

Acquisition assumptions and organization

Keep each emission line and calibration relationship explicit. Folder names do not control the workflow after import, but a stable structure helps prevent an Ha flat from silently serving OIII or SII:

project/
  lights/Ha/
  lights/OIII/
  lights/SII/
  lights/OSC/                 optional dual/tri-band branch
  calibration/bias/
  calibration/dark/<exposure>/
  calibration/flat/Ha|OIII|SII/
  calibration/dark-flat/
  broadband/RGB/              optional continuum or star branch
  output/                     keep outside source folders

Input	Match or record	Reject or isolate when
Mono lights	Filter, object, dimensions, binning, gain/offset/readout, exposure, temperature/session, time	Cloud, tracking, focus, saturation, framing, or metadata would weaken registration/normalization
OSC lights	Same capture fields plus CFA pattern and passband/filter identity	CFA is unknown, frames were pre-demosaiced inconsistently, or different dual-band filters are mixed
Bias/dark/dark-flat	Camera mode, binning, dimensions; dark exposure and temperature policy	A fallback is only dimensionally similar or would scale an unsuitable sensor pattern
Flats	Filter, optical train/orientation, focus, camera mode, binning, dimensions	Dust/focus/rotation/filter changed, or the calibration master is missing without an explicit warning decision
Broadband RGB/stars	Same target, session provenance, compatible PSF, registered/cropped geometry	Stars double, halos shift, or a different field/crop is being treated as equivalent

Image Integration does not modify source frames. It can create calibration masters, registered lights, normalization artifacts, FITS masters, previews, history snapshots, logs, manifests, and HTML/Markdown reports. Rejection maps, drizzle products, and debug-calibrated intermediates increase disk use.

Linear/nonlinear state and precision

• Calibrate, register, normalize, stack, correct gradients, and match intensity while the mono/OSC data is linear.
• OSC calibration must occur before demosaic. Do not debayer bias, dark, or flat frames just to make them display in color.
• Viewer autostretch is inspection only. Saving an image does not bake the viewer stretch into its pixels.
• Channel Composition can preserve a linear handoff. Narrowband Palette Composer and NB Color Mapper process outputs are tagged nonlinear by their process contracts; consult their pages before feeding them to a step that requires linear data.
• Keep high-precision FITS/XISF working masters. Make PNG/JPEG delivery copies from a deliberately stretched copy, not from the only linear master.

Workflow map and decision points

The release guide does not include a diagram of this workflow. The physical narrowband OSC and matched broadband-continuum branches have not passed their representative-data gate, so a combined visual flow would imply verification that the current evidence does not provide. The decision table below remains available as source-derived planning guidance; see Verification data and evidence mode for the exact boundary.

Decision	Choose this path	Check before continuing
Full Image Integration or Lite	Use full Image Integration for per-filter groups, explicit matching/settings, one shared registration group, reports, history, or incremental reruns.	Execution Plan lists the intended Ha/OIII/SII groups and output names
Shared or separate registration	Use one shared reference when all channels cover the same field and sampling. Use separate groups only when you will explicitly register resulting masters afterward.	Blink stars and edges; equal dimensions are not proof of alignment
Ha or another reference	Ha often supplies abundant structure, but choose the cleanest structurally representative frame/master.	Reference is not clouded, saturated, trailed, or near a bad edge
Direct mapping or palette process	Channel Composition for exact RGB role assignment; Palette Composer for its 12 fixed formulas and OSC synthesis; NB Color Mapper for flexible hue/saturation/layer blending.	Inputs satisfy each process's state, geometry, and OIII requirements
HOO, SHO, or custom	HOO needs Ha/OIII; SHO needs SII/Ha/OIII; custom mapping is useful when fixed matrices hide weak structures.	Histograms are not clipped and no one channel dominates only because it was left unmatched
Continuum enhancement	Use NBRGB only with a real registered RGB base and meaningful effective bandwidths for production.	RGB/NB geometry, background, units, channel assignment, and widths are verified
Direct or star-separated finish	Direct finish avoids external dependencies. Use native masks for local star protection/reduction; use StarNet only after its external prerequisites pass.	Preserve a pre-star copy and verify the recombination at 100%

Step-by-step procedure

1. Audit the acquisition set

1. Separate mono lights and flats by filter. Separate OSC raw lights and their calibration frames from already demosaiced products.
2. Compare dimensions, binning, gain/ISO, offset/readout mode, exposure, temperature/session, filter, CFA, object, and optical-train history.
3. Inventory bias, dark, dark-flat/flat-bias, and flats. A missing item becomes an explicit limitation, not an assumed match.
4. Choose an output directory outside the sources and preserve acquisitions read-only.

Checkpoint: every frame has one intended group, each flat has the correct filter, and every calibration relationship can be explained before execution.

Recovery: correct metadata or isolate incompatible frames. Do not merge groups merely to silence a missing-master warning.

2. Import, group, and validate in Image Integration

1. Open Process > Preprocess > Image Integration or its Dashboard card.
2. Add Bias, Darks, Flats, and Lights with Add Files... or Add Folder....
3. Set Output Directory. Verify that filter remains a grouping field for mono lights/flats and that exposure/session/camera-mode splits are intended.
4. On Workflow > Calibration, inspect coverage, assignments, overrides, fallback indicators, and issues.
5. Open Execution Plan. Count the expected calibration, registration, alignment, normalization, and stacking groups before running.

Checkpoint: the plan contains one output per intended emission-line group, and no frame is assigned to a master from another filter/camera mode.

Recovery: fix metadata, grouping, matching policy, or explicit assignments; rebuild and reread the plan before execution.

3. Calibrate by filter

Start with the Image Integration recommendations, then choose rejection from the actual frame count. The bounded worked run used:

Product	Combination/rejection	Normalization	Handoff
Ha/OIII/SII flat masters	Average; Sigma Clipping 3.0/3.0	Global, Additive With Scaling, IKSS; no weighting	Eight selected flats per filter; flat auto-level on during light calibration
Ha/OIII/SII lights	In-memory calibration	Per-filter flat	Cosmetic correction off; no bias/dark was available in the fixture

The worked run's three uncalibrated-flat warnings are evidence that the report exposes this weakness. For production, supply correctly matched bias/dark-flat or dark calibration where your sensor/workflow requires it and recheck the report.

For OSC, keep raw CFA data through calibration. Supply the correct CFA pattern when Auto cannot determine it, and inspect the demosaiced registered result for checkerboards, swapped colors, or zippering.

Checkpoint: every master lists its intended inputs; calibrated/registered views show no new clipping, dust inversion, residual glow, or CFA artifact.

Recovery: stop at the calibration stage, replace the wrong master or CFA setting, and rerun affected groups rather than masking the artifact later.

4. Register, align, normalize, and integrate per filter

1. Select one strong reference across compatible Ha/OIII/SII lights. The worked run used a fixed Ha light and one 24-light registration group.
2. A conservative native starting point is Star Matching, Homography, transform behavior Auto, alignment Standard, interpolation Auto. Use a simpler transform when residuals support it; use polynomial models only with enough reliable matches and genuine distortion.
3. For each light group, the worked run used Average, Winsorized Sigma Clipping 3.0/3.0, Global Additive With Scaling, IKSS, no weighting, no drizzle, and no final solve. These reproduce evidence; tune rejection/weighting from your set and the stacking reference.
4. Set maximum cores without starving the desktop; the worked run used 8.
5. Execute, monitor the timeline, and wait for a safe cancellation checkpoint if you stop the workflow.
6. Inspect the HTML/Markdown report, registered lights, normalization choices, frame retention, rejection, master previews, and warning/error section.

Checkpoint: every required group succeeds, stars share one grid, output edges are understood, expected masters exist, and each warning has an explicit fix/accept decision. success=true alone is insufficient.

Recovery: replace a weak reference or bad frames and rerun the affected registration/integration scope. Use history/last-successful state to distinguish reused artifacts from regenerated products.

5. Inspect, crop, and correct backgrounds while linear

1. Open Ha, OIII, and SII masters. Use viewer autostretch only to compare them.
2. Inspect Image Statistics and Histogram Analysis for finite data, clipped endpoints, abnormal plateaus, or nearly empty channels.
3. Crop common registration borders with Crop. Apply exactly the same crop/grid to every channel.
4. Run Background Extraction separately on each channel. Keep samples off real nebulosity and bright halos; inspect the background model for target structure.

The representative evidence crop used 20 samples/line, tolerance 2.0, RBF, smoothing 0.50, and subtraction. Use those only as an initial diagnostic.

Checkpoint: corrected channels remain finite/nonblank; the model contains smooth low-frequency background rather than recognizable nebular detail; stars and crop geometry still coincide.

Recovery: move/remove contaminated samples, reduce model flexibility, or restore the integrated master and retry. Do not repair an overfit model with later clipping.

6. Match channel intensity deliberately

Use Linear Match when a common intensity relationship is needed before direct composition. In the worked run, Ha was the reference and OIII/SII were matched with accepted range 1e-7..0.875.

1. Choose a structurally representative reference, not automatically the brightest channel.
2. Ensure source/reference channel count is equal and geometry is registered.
3. Exclude clipped shadows/highlights while keeping enough ordinary signal for a stable fit.
4. Compare before/after histograms and several background/emission/star areas.

Checkpoint: channel scale/background is more compatible without clipping, sign inversion, amplified gradients, or colored star rims after a preview composition.

Recovery: widen the accepted range only after checking endpoints; otherwise fix background/registration or choose a better reference. Preserve unmatched masters for restart.

7. Build HOO, SHO, and custom palettes

Choose the smallest tool that expresses the intended mapping:

Goal	Route and starting point	State/output caution
Exact HOO	Process > Color Spaces > Channel Composition; RGB mode; R=Ha, G=OIII, B=OIII	Enable its Linear Match only if you intend that extra matching; retain mono sources
Exact SHO	Channel Composition; R=SII, G=Ha, B=OIII	Verify common geometry and clipping before color work
Compare fixed formulas	Process > Color Spaces > Narrowband Palette Composer; direct Ha/OIII/SII roles; begin at 100% intensities	Requires OIII plus Ha or SII; process output is tagged nonlinear
Flexible custom mix	Process > Color Spaces > NB Color Mapper; one mono layer per role; set hue/saturation, then screen blend	Mapper colorization/blend is nonlinear; tune layers one at a time

Palette Composer includes SHO, HOO, HSO, HOS, OSS, OHH, OSH, OHS, HSS, Realistic½, and Foraxx. With prepared roles H, O, and S, SHO is (S,H,O) and HOO is (H,O,O). The engine clips each raw channel to [0,1] and then normalizes the entire RGB image by one global peak. Intensities change relative roles, but values already clipped at one cannot respond further.

The evidence custom map used SII hue/saturation 18/0.82, Ha 52/0.72, and OIII 195/0.78, with no output STF. It is an execution checkpoint, not a recommended aesthetic palette.

Checkpoint: output is three-channel and finite/nonblank; no channel is clipped solely because its input was left unmatched; stars do not double; the result expresses the intended structure at several zoom levels.

Recovery: return to matched mono masters, reset intensities/layers, compare HOO/SHO before complex formulas, and change one role at a time. If OIII is missing, supply it; the palette engine rejects that state.

8. Complete the narrowband OSC branch

After calibrated raw OSC data is demosaiced/integrated into a registered RGB master, Narrowband Palette Composer can synthesize roles:

Input port	Synthesized roles
OSC1	Ha from red; OIII from mean of green and blue
OSC2	SII from red; OIII from mean of green and blue

If an explicit role already exists, synthetic input is blended 50/50 with it. With explicit OIII plus both OSC ports, processing order affects the final OIII weights; read the complete composer page before mixing all paths. OSC synthesis does not know a filter's measured passbands, exposure balance, CFA leakage, or sensor response.

1. Verify calibration-before-demosaic and the CFA pattern in Image Integration.
2. Confirm what the physical filter places in red and green/blue; do not assume every dual/tri-band filter matches the port labels.
3. Start with one OSC port, no explicit roles, and 100% intensities. Inspect HOO or the appropriate two-line palette.
4. Add a direct mono role only when its registration and weighting are known.

Checkpoint: synthesized roles are plausible, geometry matches, and HOO/other outputs contain neither CFA artifacts nor unexplained channel duplication.

Recovery: return to the integrated RGB master, verify CFA/filter identity, disable one OSC helper, or extract/prepare explicit channels through a process suited to your physical filter. The worked run's same-target derivative proves synthesis arithmetic only; it does not replace physical OSC validation.

9. Add narrowband signal to RGB with continuum handling

Use NBRGB Combination only when you have a registered linear RGB target and same-size mono narrowband sources. Enter documented effective bandwidths in nanometres; AfterNight does not read measured filter curves or central wavelengths.

A conservative reproduction profile is RGB bandwidth 100, red narrowband 7/scale 1.20, green/blue 8.5/scale 1.20, fit rejection 0..0.92, and Non Linear rescale off. Replace widths with those of your actual filters. Keep every narrowband width meaningfully below the RGB width.

Tune representative Preview areas containing background, emission, ordinary stars, and highlights. Apply has no visible progress/Cancel control, so preview before full-image execution.

Checkpoint: unselected RGB channels are unchanged; selected channels add emission without exploded background noise, clipped cores, doubled stars, or colored residuals; widths and assignments are recorded.

Recovery: fix registration/background first, verify bandwidth units, reduce Scale, turn nonlinear rescale off, and widen the fit interval only when too few ordinary pixels survive. Restore the original RGB rather than stacking several failed enhancements.

10. Stretch, color-balance, and manage noise

1. Preserve the linear mono and composition masters.
2. Use Generalized Hyperbolic Transformation, Histogram Transformation, or ArcSinh Stretch for a gentle first permanent stretch.
3. Inspect faint structures, background, bright cores, and stars at 100%.
4. Shape color after the major stretch with Curves Transformation, Adjust HSL, or Color Saturation.
5. Use masks for contrast/sharpening so noisy backgrounds and halos are not enhanced with the target.

Checkpoint: black point retains faint signal, highlights are not flat, background chroma is controlled, and weak OIII/SII detail is not erased by an Ha-dominated stretch.

Recovery: restore the last linear or pre-stretch checkpoint, reduce the first stretch, and correct color/noise in smaller deliberate stages.

11. Choose and verify a star path

Three supported decisions have different prerequisites:

1. Direct finish: leave stars in the image. This is the dependency-free default.
2. Native mask-limited work: create a Star Mask, use it to protect/select stars during nonlinear operations, and compare the result with the untouched pre-mask copy.
3. Separated star/starless work: use StarNet++ only after its executable/model configuration passes. Keep matched starless and stars products, process gently, then use Starmask Recombination or the profiled variant.

Starmask Recombination performs additive recombination when no blend mask is used. Therefore the starless and stars images must come from compatible geometry/state and must not have been independently clipped into an impossible sum. Broadband RGB stars must be registered/cropped to the narrowband result and checked for PSF/halo mismatch.

Checkpoint: the star mask is neither empty nor full-frame; starless and star products are finite/nonblank; recombination has no seams, dark rings, doubled stars, clipped cores, or background shift at 100%.

Recovery: restore the pre-star image, soften/fix the mask, reduce star processing, realign broadband stars, or recombine from products made at the same stretch/state. Do not treat the worked run's native-mask split proxy as StarNet output.

12. Export and archive restart points

1. Save a high-precision working master with processing history: FITS/XISF for the native working chain or TIFF where external editing requires it.
2. Reopen it and verify geometry, channels, bit depth/pixel type, processing state/history where supported, and finite/nonblank statistics.
3. From a deliberately stretched copy, save PNG/JPEG for sharing. Reopen and inspect it at 100% and at delivery size.
4. Archive source manifests/presets, Image Integration report/history, separate Ha/OIII/SII and RGB masters, pre-stretch composition, pre-star image, star/starless products, final working master, and delivery copies.

Checkpoint: the working master and share copy both reopen; the working file retains the expected precision/state/history; the share copy looks like the intended pixels rather than the viewer STF.

Recovery: if a save fails, keep the in-memory view open, create/choose a writable parent, verify free space and extension, retry, then reopen. The worked run verified this missing-parent failure/retry path.

Parameter handoff matrix

From step	Output/checkpoint	Decision criterion	To step	Parameters/state carried forward	Recovery
Acquisition audit	Classified frame inventory	Filter/camera/session/calibration relationships are explicit	Image Integration	Metadata, source paths, output root	Correct/isolate frames before import
Calibration	Per-filter calibrated lights/masters	No wrong-filter flat, CFA artifact, clipping, or unexplained warning	Registration	Linear state, filter role, camera mode	Replace master/CFA setting and rerun affected scope
Registration/integration	Ha/OIII/SII or OSC masters plus report	Required groups succeed; common grid and warnings accepted	Background preparation	Geometry, reference provenance, normalization/stack settings	Change reference/bad frames and rerun
Background preparation	Cropped/corrected channels	Model contains background, not target; grids remain identical	Linear Match	Linear state, common crop, channel identity	Restore master and refit samples/model
Linear Match	Reference-scaled mono roles	No clipping, inversion, gradient amplification, or halo shift	Palette branch	Reference identity and accepted range	Change range/reference or fix upstream
Mono roles	HOO/SHO/Foraxx/custom result	Three-channel finite/nonblank result expresses intended roles	Nonlinear finish	Formula, role intensities/hues, process state	Reset to matched masters and simpler mapping
OSC master	OSC1/OSC2 synthesized palette	Physical passband-to-role assumption is verified	Nonlinear finish	CFA/filter identity, chosen helper, intensities	Disable helper or rebuild explicit roles
RGB + NB roles	NBRGB result	Widths/registration/background recorded; enhancement is controlled	Nonlinear finish	Bandwidths, scales, rejection, rescale policy	Restore RGB; fix upstream; reduce Scale
Nonlinear finish	Pre-star final candidate	Stretch/color/noise checkpoints pass	Direct export or star path	Nonlinear state, masks, untouched checkpoint	Restore pre-stretch/pre-star checkpoint
Star path	Verified recombination	Geometry/state match; no rings/seams/doubling	Export	Star method, mask, recombination settings	Rebuild products from same source/state
Export	Reopened working/share files	Precision, history/state, geometry, and visible content pass	Archive/publish	Hashes, filenames, report/preset/history	Keep view open, repair path/space, retry

Algorithms and why the order matters

Calibration, registration, and matching

Calibration corrects detector/optical patterns before interpolation. Demosaicing OSC data first changes the sampling and prevents raw calibration masters from operating on the original CFA pixels. Registration resamples every role onto a common grid; palette math cannot repair translation, scale, rotation, distortion, crop-origin, or PSF mismatch.

Linear Match estimates an affine intensity relationship over accepted samples. For equal dimensions it fits paired pixels; for different dimensions it uses independent moments, which cannot establish spatial alignment. Background correction therefore precedes matching, and registration precedes both.

Palette math and process state

For prepared Ha/OIII/SII samples H, O, S:

\[ SHO=(S,H,O),\qquad HOO=(H,O,O). \]

Palette Composer also implements weighted Realistic and nonlinear Foraxx formulas. Inputs are clipped, formulas are evaluated, and a global RGB peak normalizes the output. This makes palette intensities relative rather than a simple exposure-preserving scale. Its optional input autostretch is a nonlinear preparation; applying it to already stretched sources double-stretches them.

NB Color Mapper maps each mono layer through hue/saturation and tonal controls, then combines included colorized layers with an N-layer screen blend. It is a creative nonlinear composition, not a spectral reconstruction.

OSC synthesis

OSC1 uses red as synthetic Ha and mean green/blue as synthetic OIII. OSC2 uses red as synthetic SII and the same mean as synthetic OIII. Explicit and synthetic roles mix at fixed weights; no filter transmission, exposure, or noise model sets those weights. Verify the physical filter before accepting the labels.

Continuum arithmetic

For selected channel c, NBRGB treats the RGB and narrowband filters as rectangular passbands. With broadband sample R_c, narrowband sample N_c, RGB width B_rgb, and narrowband width B_c:

\[ q_c=\frac{B_c}{B_{rgb}},\qquad C_c=\frac{N_c-q_cR_c}{1-q_c}. \]

It median-centers and scales C_c, attempts a guarded affine fit over the accepted range, and takes a per-pixel maximum with the boosted RGB plane. Because 1-q_c approaches zero when widths approach equality, wrong bandwidths can amplify noise severely. The equation ignores measured transmission, central wavelength, exposure, sensor response, and spectra; use it as an approximation.

Stars, masks, and export

Star operations follow the major stretch only when the chosen process expects nonlinear data. A mask must be checked for coverage before it controls a strong operation. Additive recombination assumes compatible decomposed products; independent clipping breaks that identity. Saving comes last, but restart masters are saved throughout: irreversible color/star experiments should never replace the only linear channel masters.

Performance and resource planning

• Image Integration cost grows with pixels, frames, registration transforms, rejection, output maps, and drizzle. Registered lights alone can approach the selected light-set size; history and debug products add more.
• Palette Composer's grid calculates all 12 full-resolution float RGB results synchronously for preview and has no user-facing cancellation. Tune on smaller representative data where possible, then wait for completion.
• NB Color Mapper and NBRGB allocate several full-size intermediates. NBRGB processes selected channels serially while pixel loops use oneTBB; nonlinear rescale adds full-plane scans.
• Debug validation is not a user-performance benchmark: the same bounded integration of the worked run took 20.451 seconds in Release and 777.882 seconds in Debug.
• Use representative crops for parameter discovery, then confirm one full-frame preview before Apply. A crop cannot prove full-frame gradients/highlights.
• Cancellation in Image Integration is cooperative. Wait for the safe stage boundary, inspect partial/history state, and confirm what will be reused before resuming.
• No network/catalog/external tool was required for the validated native path.

Verified worked run

• Sample-set provenance: the mono SHO data set of M16 captured on 2025-06-18, project-owned and supplied by the maintainer; 48 deterministically selected inputs, all hashed under sanitized IDs.
• Application platform: Windows MSVC x64 Release and Debug, Qt 6.10.3.
• Integration: eight 300-second lights and eight flats per Ha/OIII/SII; fixed Ha registration reference; 24/24 effective and registered; three 3008 x 3004 32-bit float FITS masters; zero errors; three expected warnings for flats without bias/dark calibration.
• Downstream crop: centered 1024 x 1024; background sample counts Ha 323, OIII 357, SII 371; OIII/SII matched to Ha at 1e-7..0.875.
• Palettes: direct SHO, HOO, and Foraxx plus custom NB Color Mapper all finite/nonblank. Missing OIII raised the expected process error.
• OSC boundary: a same-target RGB derivative exercised OSC1 red-to-Ha and mean-green/blue-to-OIII; derived HOO matched direct HOO within sampled mean absolute delta 1e-6. No physical OSC claim is made.
• Continuum boundary: a same-target mapped RGB derivative exercised NBRGB rectangular-bandwidth arithmetic and changed the image. No measured broadband-continuum or photometric claim is made.
• Stars: native Star Mask split proxy selected 0.744%; starless/stars were finite/nonblank; additive recombination matched the stretched input within sampled mean absolute delta 1e-6. No StarNet execution claim is made.
• Export/recovery: the historical 1024 x 1024 x 3 32-bit float FITS reopened with a linear state that the old test assigned incorrectly. The corrected harness follows the production Palette Composer nonlinear contract and awaits an approved-data rerun. The historical HISTORY and finite/nonblank content checks remain valid; the 16-bit PNG reopened finite/nonblank, and the missing-parent save failed then passed after parent creation and retry.
• Regressions: the current remediation Release and Debug runs each passed 61 focused engine and 21 focused process/UI contract tests. The corrected private-data filters compiled and skipped 2/2 in both builds because the approved-data variables were absent; a fresh approved-data run remains required. The standard repository-wide Release runner passed all discovered Test_*.exe test executables in 1,067.6 seconds.

Hashes, exact settings, output observations, privacy policy, and exclusions are in the guide's validation record.

Warnings and limitations

Verification remains blocked on representative branch data

The physical mono branch has partial execution evidence, but no approved physical narrowband OSC or matched broadband-continuum set was supplied. Same-target arithmetic proxies do not satisfy the every-branch mode-E gate. The corrected palette-state harness must also be rerun with approved data.

Calibrate flats correctly for production

The supplied session had per-filter flats but no approved bias/dark data. The run intentionally retained three uncalibrated-flat warnings as evidence. Do not copy that limitation into a production workflow.

The OSC and continuum executions are bounded proxies

They verify current native algorithm paths on representative same-target signal. They do not verify raw CFA/demosaic/sensor behavior, physical dual-band separation, measured continuum, filter transmission, or photometric fidelity.

Equal dimensions are not registration

Palette and continuum tools generally validate dimensions/types, not WCS, star alignment, crop origin, PSF, or filter identity. Misregistration can produce plausible-looking colored halos.

Palette tools are not spectrophotometry

Fixed formulas, global normalization, HSV colorization, screen blend, and user-entered bandwidths create mapped color. They do not recover physical emission ratios without a separate calibrated measurement workflow.

Optional star tools require separate readiness

The worked run did not use StarNet. Configure and validate its executable/model before relying on a starless branch; keep a direct native finish available.

• The worked values describe one bounded sample and representative crop.
• No full-frame polished result, print, Linux, installer, catalog, solver, network, external model, or general performance claim is made.
• Process pages remain authoritative for complete controls, state behavior, masks, progress/cancellation gaps, and metadata limitations.

Troubleshooting and recovery

Stage/symptom	Verified cause or failed checkpoint	Recovery	Resume/restart consequence	Evidence
Flat warning	No bias/dark master assigned to that filter's flat group	Supply a compatible master or explicitly accept the bounded limitation	Rebuild flat and downstream light groups	Worked run: three warnings
Expected filter master missing	Wrong grouping/filter label, failed stage, or output collision	Correct metadata/plan/output name and rerun affected group	Preserve successful unrelated groups/history	Image Integration source/run
Low-signal OIII/SII rejected	Weak frames, normalization/reference/rejection mismatch	Inspect frame metrics; change reference/rejection or exclude only verified failures	Rerun affected normalization/stack	Workflow report contract
Colored halos/doubled stars	Channels have unequal registration/crop/PSF	Register/crop on one grid; inspect stars at center/edges	Restart from mono/RGB masters	Source + decision checkpoint
One role dominates palette	Background/intensity mismatch or clipped input	Correct backgrounds, Linear Match deliberately, reduce intensity/layer saturation	Restart composition from preserved masters	Worked-run matching/palettes
Palette Composer rejects inputs	OIII absent, no Ha/SII after synthesis, invalid/missing image, or unresolved sizes	Supply required role, fix/enable size preparation, reopen valid views	No output; inputs unchanged	Negative path + process page
OSC palette looks implausible	Physical passbands/CFA do not match OSC helper assumptions	Verify filter/CFA; disable helper; prepare explicit roles	Restart from calibrated OSC master	Scoped OSC execution
NBRGB background/noise explodes	Bad registration/background, wrong units/widths, or width near RGB width	Fix upstream; verify widths; reduce Scale; keep nonlinear rescale off	Restart from original RGB	Source + scoped arithmetic
NBRGB fit seems inactive	Fewer than three usable pairs or near-constant corrected plane	Inspect clipping and rejection; widen range carefully	Fallback uses corrected NB directly	NBRGB source/tests
Star mask empty/full	Detection threshold/size/shape unsuitable	Inspect mask, tune one control, regenerate before processing	Restore pre-star checkpoint	Worked-run native star path
Recombination has seams/rings	Mask too hard or star/starless states no longer form a compatible pair	Soften/rebuild mask; recombine matched products	Restart from pre-star image	Recombination source/checkpoint
Saved file is dark	Only viewer autostretch was visible	Permanently stretch a copy, then export it	Linear master remains valid	Save/viewer contract
Save fails	Missing/unwritable parent, disk, lock, or extension issue	Keep view open; repair destination; retry; reopen	No processing restart required	Worked-run failure/recovery

Related workflows and references

• Image Integration — calibration, common registration, per-filter integration, artifacts, and reports.
• Linear Match — accepted range, affine fit behavior, geometry, and state.
• Channel Composition — exact HOO/SHO assignments and optional linear matching.
• Narrowband Palette Composer — OSC roles, formulas, intensities, global normalization, UI execution, state, and provenance.
• NB Color Mapper — custom layer colorization, modifications, screen blend, and output behavior.
• NBRGB Combination — continuum equation, bandwidths, fit/merge, performance, and limitations.
• Star Mask, StarNet++, and Starmask Recombination — direct/native/external star paths.
• Saving and Exporting and File Formats — working masters, delivery copies, processing history, and readback.
• Broadband Imaging — production OSC and RGB/LRGB acquisition/integration context for real continuum or broadband-star inputs.
• The guide's validation record — the sanitized execution record and claim boundaries.

The palette-composer and mapper process pages record their upstream source-code provenance and license constraints. NBRGB is native AfterNight GPL code. This workflow adds no external algorithm, asset, model, or network dependency.