Calibration Foundations

What calibration is, and what it is not

Calibration corrects the detector and the optical path against measured references before any creative processing begins. It is not an appearance adjustment, and it is not something a later step can stand in for. The distinction matters because several later operations can produce a superficially similar before-and-after while being physically different corrections.

One category error is worth naming up front: bias, dark, and pedestal terms are additive; flats are multiplicative; debayer is an interpolation and a change of state; cosmetic correction replaces localised defects; and background extraction is a model derived from the science image itself. Treating any of these as interchangeable is how a calibration problem gets papered over instead of fixed.

Bias and electronic offset

A detector readout carries an electronic baseline in addition to photon-derived signal. A bias or equivalent short-exposure reference characterises that baseline, though some CMOS workflows fold it into darks or flat-darks instead of keeping a separate master bias.

The public distinction that matters most: bias is an acquisition property of the sensor, while image background is sky plus terrestrial light plus detector noise plus whatever processing has already happened. A raised background is therefore never, by itself, an identification of an electronic offset.

Dark calibration

Dark frames capture the exposure-, temperature-, and gain-dependent fixed signal -- dark current and hot-pixel structure -- along with readout contributions. Whether subtraction is correct depends on how the masters were built and whether dark scaling or optimisation is valid for that particular camera.

For SeeStar data specifically, ZWO has stated that exported files are dark calibrated. That is a real provenance fact worth relying on, and it stops short of establishing that the same files received independent bias and flat calibration.

Flat calibration

Flats correct multiplicative variation in sensitivity and illumination: vignetting, pixel-response nonuniformity, dust shadows, and optical differences across the field. The algebra depends on which masters already contain the additive terms, so the roles matter more than any single universal formula.

Flat correction and background extraction are not substitutes for one another. A flat is a measured correction from a reference exposure; background extraction is a model fitted to the science image. The second can reduce how residual vignetting looks without making the data correctly flat-fielded.

CFA state and debayer order

For one-shot-colour data, calibration should normally happen while the frames are still in their colour-filter-array mosaic state. Demosaicing interpolates the missing colour samples, after which neighbouring pixels no longer represent independent raw sensor readings in the same way.

The robust teaching order is raw CFA, then calibration, then debayer, then registration, then integration. Engines may fuse those into fewer commands or save different intermediates, but a wrong Bayer pattern or a debayer performed at the wrong stage is not something later colour work can undo.

Clipping and output pedestals

Calibration arithmetic can legitimately produce values below zero. If the software or file representation clamps those to zero too early, the information is gone: a clipped zero does not remember how far below zero it would have been. This is why an output pedestal can be added during calibration -- a known positive offset that keeps legitimate low samples representable through truncation.

That pedestal is bookkeeping, not signal, and the consequence runs one way only. If calibration has already clipped, adding an offset afterwards raises zeros without restoring what separated them. Choose the amount from measured clipping diagnostics for the data and tool at hand; published numeric examples are procedural guidance for a specific tool and version, not a constant.

Input pedestal versus output pedestal

Two different things share the word pedestal. A calibration-time output pedestal is protection added on purpose so that low values survive truncation. A residual electronic offset is an uncorrected additive term left in the data. They are different claims about different stages of the same pipeline.

Confusing them has a specific failure mode: subtracting a protective pedestal later, on the theory that it is a residual offset, can reintroduce exactly the clipping it was added to prevent. The narrower Pedestal Removal article covers when a genuine residual offset may be removed -- and its default answer remains skip.

What SeeStar has already done

Exported SeeStar files arrive with vendor-stated dark calibration. Treat that as a provenance attribute rather than a guarantee: it is version-bound, it does not assert independent bias and flat correction, and community reports of clipping in some firmware-era files mean that low-end damage is worth diagnosing rather than assuming away.

A native SeeStar stack is already far downstream of raw calibration. Redoing a dark or flat calibration on it is not possible without the original subframes and appropriate masters.

What NOVA can still do after stacking

NOVA's processing pipeline starts from a stack, which means the CFA and calibration decisions have already been made by the capture or stacking path before the first processing step runs. What remains available is detection and mitigation of residual defects, not reconstruction of raw measurements.

Stated plainly: cosmetic correction can suppress leftover isolated defects, background extraction can model residual large-scale gradients, colour calibration can correct channel response, and a documented residual offset can be subtracted. None of those recreate correctly calibrated pixels, and describing them as repairing calibration would overstate what they do.

Is this a calibration problem, or something else?

Persistent hot or cold pixels in stable sensor positions, amp glow, and structured dark residuals point upstream to dark calibration. Vignetting, dust shadows, inverse dust donuts, and repeated multiplicative field patterns point to flats. Zipper or checker patterns, wrong colour assignment, and altered star profiles point to debayer placement or pattern. A sharp pile-up at zero points to calibration clipping.

The corresponding recovery in each case is to return to the original subframes and recalibrate, not to reach for a downstream tool. Cosmetic correction, background extraction, and colour calibration each have their own article and their own legitimate job; standing in for absent calibration is not it.

How to validate calibration

Measure on the calibrated subframes, before later processing can hide the difference: the low-clipped fraction with known file scaling, per-channel or per-CFA-plane minima, medians, robust spread and clipping counts compared at the same processing state, flat-field residual spatial structure, and persistent hot or cold pixel structure across frames. A difference image against a controlled reference shows whether a claimed scalar operation is actually spatially uniform.

Several tempting metrics cannot carry this weight. A background-gradient score is confounded by real nebulosity and by later modelling; a whole-image signal-to-noise proxy is not photometric signal-to-noise and rewards smoothing and rejection; star count, half-flux diameter, and eccentricity are dominated by seeing, registration, and extraction thresholds. Visual inspection is genuinely useful for gradients, dust donuts, clipped black texture and mosaicing artefacts -- as a supplement to subframe statistics, not as a replacement.

Comparing calibration strategies fairly

A calibration comparison has to happen upstream on the same raw subframes, not by subtracting different constants from an image that has already been stacked. Fairness requires identical frames and inclusion sets, matched masters or an explicitly stated master-strategy difference, identical CFA and debayer ordering, identical registration and integration settings, identical bit depth and output scaling, and an identical preview stretch used only for review.

The confounders that most often invalidate such a comparison are different frame sets, differently built masters, different demosaic methods, different rejection or weighting during integration, an unnamed difference in output pedestal or clipping policy, firmware drift between captures, and adaptive downstream steps that retune themselves to the changed statistics. Any result from such a comparison is a result for that data, those candidates, those versions, and that evaluation procedure.

Evidence and version notes

Calibration behaviour is engine- and firmware-dependent, so the strength of each statement here varies deliberately. Siril's support for bias, dark, and flat masters with CFA-aware handling is vendor documented. PixInsight's calibration model is vendor documented, while the output-pedestal practice is drawn from vendor community guidance and its numeric examples are procedural rather than specification. The SeeStar dark-calibration statement is direct but narrow vendor evidence.

Two claims are explicitly not supported and should not be repeated: that SeeStar calibration can never clip, and that a flatter final background demonstrates improved flat calibration. Neither follows from the available evidence.

Auditable claims

1. Calibration and post-stack correction are different layers. Bias, dark, and flat calibration act upstream on raw sensor frames; a post-stack residual-offset step is a narrow scalar correction and is not a substitute for correct calibration.

   source-confirmed · none

2. Once calibration has clipped low values to zero, adding a pedestal later cannot reconstruct the lost values or the differences between them. It only raises zeros that have already lost their original spacing.

   source-confirmed · none

3. A positive minimum, a positive median, or a raised background is not enough to identify an electronic offset. Image background is sky, terrestrial light, detector noise, and processing state combined; bias is an acquisition property of the detector readout.

   source-confirmed · none

4. Flat correction is a multiplicative correction measured from an optical and sensor reference. Background extraction is a model inferred from the science image itself. A gradient tool can make residual vignetting look better without being calibration-equivalent to a proper flat.

   source-confirmed · none

5. For one-shot-colour data the conceptual order is raw CFA, then calibration, then debayer, then registration, then integration. Demosaicing interpolates missing colour samples and changes the pixel statistics, so calibrating afterwards is no longer acting on independent raw sensor samples.

   vendor-documented · Individual engines may fuse commands or save intermediates differently; the ordering is conceptual, not one fixed command sequence. · none

6. An output pedestal added during calibration is bookkeeping that keeps legitimate low-valued samples representable before truncation. It is protection against destructive clipping, not added astrophysical signal.

   vendor-documented · Published numeric examples are tool-, version-, and data-specific procedural guidance; choose the amount from measured clipping diagnostics rather than adopting a universal default. · none

7. A protective calibration-time output pedestal and a residual electronic offset in a finished stack are different claims about different stages. Subtracting a protective pedestal later can recreate exactly the truncation it was added to prevent.

   source-confirmed · none

8. ZWO states that FITS exported by SeeStar are dark calibrated. That is a useful provenance fact and does not establish that every exported frame also received independent bias and flat calibration.

   vendor-documented · Calibration provenance is version-bound: firmware and capture-app behaviour can change what a given file actually received. · none

9. A file being produced by SeeStar does not make low-end clipping impossible. Community incident reports describe clipping and black-level problems in some firmware-era calibrated files, so suspicious clipping is worth diagnosing rather than ruling out.

   vendor-documented · Incident evidence, not a general specification: it establishes that version-specific defects can occur, not that any particular capture is affected. · none

10. NOVA's post-stack pipeline begins from an already-integrated image, so CFA-domain calibration and debayer decisions have already been made upstream by the capture or stacking path. The post-stack pipeline cannot reconstruct raw-frame calibration from the stack alone.

    source-confirmed · none

11. Cosmetic correction can suppress isolated residual defects, background extraction can model large-scale residual gradients, and colour calibration can correct channel response -- but none of them recreate improperly calibrated pixels. The honest description of downstream cleanup is that it mitigates visible residuals, not that it repairs calibration.

    source-confirmed · none

12. A flatter final background does not prove a better flat, lower noise does not prove better dark calibration, and colour balance is not a direct flat-quality metric. Each of those can be produced by later modelling, rejection, or smoothing rather than by calibration.

    source-confirmed · none

13. A fair calibration comparison starts from identical raw subframes with an identical inclusion set, matched masters, identical CFA and debayer ordering, and identical registration and integration settings afterwards. Comparing finished images from different frame sets or different downstream settings measures the whole strategy, not calibration.

    source-confirmed · none

14. The strongest acceptance evidence for calibration is measured on the calibrated subframes -- clipped fraction, structured residuals, CFA and debayer state -- before any later adaptive processing can hide the difference. Which final rendering scores higher is not the test.

    source-confirmed · none