Wavelet Denoise

Purpose and expected result

Wavelet Denoise reduces low-amplitude structure by decomposing the image into decimated Haar-wavelet coefficients, soft-thresholding them relative to a robust noise estimate, reconstructing the image, and averaging several randomly shifted repetitions. It is useful when you want direct scale control rather than a patch-trained or learned denoiser.

The expected result is a same-size, same-type still or sequence with less visible fine noise. It is not a detail-layer editor: one Threshold is applied across the entire transformed coefficient array, including the coarsest approximation. Excessive Threshold or Scales can therefore suppress real structure, create block or boundary artifacts, and alter broad levels. Every execution draws the same fixed-seed set of cycle-spinning shifts, so the same input and settings give the same result.

At a glance

Item	Current behavior
Menu path	Process > Noise Reduction > Wavelet Denoise Filter
Enabled when	An image or sequence is active, the same process is not already attached to it, and no attached process holds exclusive access
Inputs	One intended normalized grayscale or RGB still or sequence; a still CFA mosaic is treated as grayscale samples unless debayered first; an optional active mask must match target geometry
Outputs	Destructively replaces a still or, after successful temporary processing, a sequence; Take snapshot creates an auxiliary still; dimensions and pixel type are retained; component modes reconstruct RGB; no file is written directly
Recommended image state	Usually linear or early nonlinear, after calibration/registration/stacking and before a strong stretch amplifies noise; valid on nonlinear data when judged in that state
Preview support	Asynchronous RT preview using the same G'MIC command and Amount/mask blend as Apply; a selectable preview area and display-only autostretch are available
Sequence support	Supported by the generic temporary replacement-sequence path with per-frame processing, progress, and cooperative cancellation; configured CFA auto-debayer can produce RGB frames
Mask support	Supported for preview, still Apply, and sequence Apply by post-filter blending; white applies, black protects, gray blends, and Invert Mask reverses weights
External dependencies	Bundled G'MIC 3.5.4 command denoise_haar, run through AfterNight's optional G'MIC adapter; no separate installation, model, catalog, or network service. In a build without G'MIC, Preview and Apply are disabled and their tooltip says why
Non-finite samples	An image holding a NaN or infinite sample is refused before G'MIC sees it, and left unchanged

Before you start

• Finish calibration, registration, stacking, gradient correction, and color calibration first. Wavelet thresholding cannot distinguish faint signal from coefficients at the estimated noise scale.
• Work on normalized finite samples. The image is handed to G'MIC as float fractions of full scale and the result is clamped to [0,1]; out-of-range float data is therefore clipped even with a weak filter. An image holding a NaN or infinite sample is refused with "the image holds a sample that is not a finite number, which G'MIC cannot process" and left unchanged.
• This process needs G'MIC, which AfterNight normally includes. In a build without it, Preview and Apply are disabled and their tooltip reads "This build of AfterNight carries no G'MIC library, so this process cannot run."
• Start with RGB only if the same absolute coefficient threshold is appropriate for every channel. For chroma-heavy noise, compare a lightness/chroma component on a duplicate and inspect reconstructed color.
• Do not treat Iterations as convergence. It is the number of independent random cycle-spinning reconstructions averaged together. The filter has no convergence tolerance or “converged” status.
• Keep Scales proportionate to the target. G'MIC temporarily extends image dimensions to multiples required by the Haar depth; high scales can add large work buffers and stronger edge sensitivity.
• Duplicate a still or rely on a verified Image History snapshot. Apply replaces the target only after successful completion; sequence Apply uses temporary disk storage and commits only after every frame succeeds.

Procedure

1. Select a calibrated still or sequence and open Process > Noise Reduction > Wavelet Denoise Filter. Confirm Target Image, choose RGB or a deliberate component, and keep Preview selected.
2. Leave Threshold: 1.40, Scales: 6, Iterations: 10, and Amount: 1.00 for the first preview. Use Preview area around both smooth background and real fine structure; the area affects Preview and snapshots, not Apply.
3. Lower Threshold if stars, filaments, texture, or local contrast soften. Raise it only until random grain falls without making small features waxy or blocky.
4. Reduce Scales if larger structures or boundaries move. Raise Scales only when noise remains at broader spatial support and the memory/padding cost is acceptable.
5. Use Iterations to average more shifted reconstructions, not to demand more smoothing. Use Amount for the final source/result blend; values below 1 restore original signal and noise uniformly.
6. Toggle Switch image with the same Linear/Autostretch display mode and inspect at 100%. Select Apply only after checking edges, stars, color, and a smooth background. For a sequence, monitor progress; Cancel preserves the original.

Complete parameter and action reference

UI label/action	Type and unit	Runtime default	Persisted-default exception	Valid values, step, or constraints	Effect of increasing, decreasing, or selecting	Dependencies and enable/visibility rules	Quality, artifact, runtime, memory, or disk trade-off	Scope: preview/apply, channel, mask, still/sequence	Persistence
Threshold	Double value plus slider, estimated-noise multiplier	1.40	None	0.00..10.00; two decimals; spin step 0.01; slider linear over 1,000 positions	Sets soft threshold T = Threshold × sigma; larger values zero/shrink more coefficients	Enabled while idle; Apply remains enabled at 0	Higher values remove more noise and real low-amplitude detail; very high values can flatten levels because all coefficients are thresholded	Preview/Apply; selected channel/component; before Amount and mask; still/sequence	Captured per execution; Reset returns 1.40; successful Apply retains the current value
Scales	Integer value plus slider, Haar levels	6	None	2..10, step 1; actual level is capped by image dimensions	Adds coarser Haar levels and expands the largest affected support	Enabled while idle; no Auto (0) choice is exposed	More levels add transform work, can increase temporary padding sharply, and expose broader structure/boundaries to thresholding	Preview/Apply; selected channel/component; still/sequence	Captured per execution; Reset returns 6; successful Apply retains the current value
Iterations	Integer value plus slider, cycle-spinning repetitions	10	The label/tooltip says denoising iterations, but the bridge passes it as G'MIC cycle_spinning	1..32, step 1	Averages more independent randomly shifted denoise/reconstruction results; it is not an optimization loop or convergence count	Enabled while idle	Cost scales approximately linearly; more cycles reduce shift dependence but do not guarantee a stronger or identical result	Preview/Apply; selected channel/component; still/sequence	Captured per execution; Reset returns 10; successful Apply retains the current value; random shifts are not persisted
Amount	Double value plus slider, blend fraction	1.00	None	0.00..1.00; two decimals; spin step 0.01; slider linear over 1,000 positions	Blends source × (1-Amount) + denoised × Amount; 0 restores source, 1 uses full denoised result	Enabled while idle; Apply remains enabled at 0	Lower values recover signal and noise together; adds one full-image blend whenever not 1	Preview/Apply after channel reconstruction and before mask; still/sequence	Captured per execution; Reset returns 1.00; successful Apply retains the current value
Target Image	Image selector	Attached/active target	None	Open eligible image or sequence	Changes source, Channel list, Preview, mask, and Apply destination	Active image required; duplicate/exclusive gates apply	Wrong target is destructively replaced after success	Preview/Apply source and active mask	Target lifecycle; Reset does not change it
Channel	Combo box	RGB for RGB; disabled Grayscale for mono	None	RGB, Red, Green, Blue, CIE L* (Lightness), CIE a, CIE b, CIE c, CIE h, CIE X, CIE Y (Luminance), CIE Z, HSL Hue, HSL Saturation, HSL Lightness, HSV Hue, HSV Saturation, HSV Value, YCbCr Luma, YCbCr Blue Chroma, YCbCr Red Chroma	RGB sends all stored channels to G'MIC with one shared noise estimate; stored/component choices filter one plane and reconstruct/copy the others	Component choices require RGB; mono selector disabled	RGB channels are thresholded independently but share one absolute threshold; converted modes add buffers, quantization, and gamut/clipping	Preview/Apply; active mask post-blend; still/sequence	Captured by current selection; Reset does not restore RGB; successful Apply retains the selection
Preview	Checkbox	Checked	None	Checked/unchecked	Enables/disables asynchronous full-image or selected-area execution	Process idle	Runs the full G'MIC filter; repeated previews cost time	Preview only; mask-aware; current still/current displayed sequence frame	Window lifetime; Reset does not change it
Preview pane	Image viewer	Current result	None	Pan with scroll bars/dragging; toolbar actions are listed separately below	Shows the mask-aware result without modifying target	Exists only with Preview	Holds source/result and display buffers	Preview only	Window lifetime
Fit To Window	Preview-toolbar action	Preview is fitted when its result is first displayed	None	Click or Ctrl+Shift+F	Sets one uniform scale that fits the whole image inside the viewport	Exists only with Preview and a displayed image	Best overview; small details may be resampled or hidden	Preview display only; no effect on Apply/snapshot pixels	Viewer lifetime; later zoom actions replace the fitted scale
Zoom In	Preview-toolbar action	Current fitted/display scale	None	Click; each activation requests 1.25× relative zoom within viewer limits	Enlarges the image around the viewport center	Exists only with Preview and a displayed image	Improves local inspection but increases panning and can expose display interpolation	Preview display only	Viewer lifetime
Zoom Out	Preview-toolbar action	Current fitted/display scale	None	Click; each activation requests 0.8× relative zoom within viewer limits	Reduces the image around the viewport center	Exists only with Preview and a displayed image	Shows more context while hiding fine residuals	Preview display only	Viewer lifetime
Zoom 100%	Preview-toolbar action	Current fitted/display scale	None	Click; sets absolute scale 1.0	Displays one image pixel at the viewer's 100% scale	Exists only with Preview and a displayed image	Preferred for judging noise, stars, and artifacts; large images require panning	Preview display only	Viewer lifetime
View Channel	Preview-toolbar combo box	RGB for RGB; disabled Grayscale for mono	None	RGB, Red, Green, Blue, CIE L* (Lightness), CIE a, CIE b, CIE a=R / b=G (Chrominance), CIE c, CIE h, CIE X/Y/Z, HSL Hue/Saturation/Lightness, HSV Hue/Saturation/Value, HSI Hue/Saturation/Intensity, and YCbCr Luma/Blue Chroma/Red Chroma	Chooses only which stored or derived component the preview viewer displays	Distinct from process Channel; RGB target required for component choices; mono is disabled	Derived display modes add conversion work and can exaggerate or hide artifacts without changing pixels	Preview display only; no effect on execution, Apply, masks, or snapshots	Preserved while the viewer rebuilds the same channel list; not in process history
Display Mode	Preview-toolbar combo box	Inherits target Linear or Autostretch mode when preview is created	None	Linear, Autostretch	Changes display only; Autostretch uses cached full-target source statistics	Exists only with Preview; Autostretch enables Autostretch Settings	Can reveal faint residuals or exaggerate noise; never changes output samples	Preview display only; no effect on Apply/snapshot	Preview-view state; not in process history
Autostretch Settings	Preview-toolbar dialog action	Disabled in Linear mode; inherits current target-view settings	Shared-view values load when opened	Opens Target Background, Shadows Clipping, linked channels, neutralization/color calibration, saturation/background, brightness, contrast, banner, Reset, and Close; Apply Stretch is hidden	Tunes only preview display	Autostretch mode only; disabled while its dialog is open	Aggressive display settings can misrepresent denoise strength/color	Preview display only	Viewer-scoped; not captured in Wavelet parameters
Switch image	Preview-toolbar action	Denoised preview shown	None	Click; Space; right-click; Ctrl+Shift+O	Toggles between current result and original target	Exists only with Preview	Comparison only; no recomputation	Preview display; still/current displayed sequence frame	Preview-window lifetime
Preview area	Preview-toolbar checkbox	Unchecked	None	Checked/unchecked; move/resize overlay	Processes a selected area and composites it over the original full preview	Exists only with Preview	Reduces transform work while tuning; final Apply remains full-frame	Preview and Take snapshot only; mask cropped consistently	Preview-window lifetime; Reset does not change it
Take snapshot	Preview-toolbar action	No snapshot	None	Click or Ctrl+Shift+K	Re-executes current settings and opens <target> - Wavelet Denoise - SNAPSHOT	Exists only with Preview	Allocates/processes an auxiliary still; display autostretch is not baked in	New still; target unchanged; selected area/mask honored	Each action creates a separate image view
Apply	Button/action	Enabled	None	One action, including Threshold/Amount 0	Captures all four values and processes selected channel for still or every sequence frame	Target must lock; controls disabled during commit	G'MIC runs serially per call; high scales/cycles and conversions dominate; mask/Amount add blends	Destructive still/sequence replacement after success	Records partial history and retains the current dialog controls after success
Reset	Button/action	Enabled while idle	None	One action	Restores Threshold 1.40, Scales 6, Iterations 10, Amount 1.00	Does not change Target, Channel, Preview, preview toolbar, or display	No output change; checked Preview schedules another execution	Configuration only	Immediate
Close	Button/action	Enabled while idle	None	One action	Closes without applying current settings	Disabled during Apply	Active preview cancellation reaches the running command through G'MIC's abort flag; pending preview discarded	Window/execution	No process settings persist after destruction
Sequence progress / Cancel	Shared indeterminate per-process/frame progress and Cancel	Created automatically for sequence Apply	None	Cooperative Cancel; frame text updates through temporary sequence processing	Writes processed frames to a temporary sequence and replaces target only after success	Sequence target only	Decode/encode, temporary disk, and repeated G'MIC execution add cost; cancellation can interrupt the active G'MIC command	Sequence Apply; selected channel, Amount, and mask apply to every frame	Cancel removes temporary output; success replaces sequence and records one history step

How the algorithm works

Conceptual pipeline

1. The selected stored image or converted component is checked for non-finite samples and handed to G'MIC through AfterNight's adapter, one G'MIC session per run, as planar single-precision fractions of full scale: a tightly packed float image in place, anything else as a working copy.
2. G'MIC estimates one noise standard deviation, chooses the requested Haar depth capped by image size, and temporarily extends dimensions to multiples of 2^(scales+1) using no interpolation and a zero (Dirichlet) boundary.
3. For every cycle, G'MIC periodically shifts the working image by a random integer displacement, drawn from a generator AfterNight seeds with the same fixed value on every call, performs the multiscale Haar transform, soft-thresholds coefficients, performs the inverse transform, undoes the shift, and accumulates the reconstruction. The accumulated image is divided by the cycle count and restored to original dimensions.
4. AfterNight clamps every output sample to [0,1], converts back to the source pixel type (rounded to nearest and held to the type's range), then reconstructs the chosen color component if necessary, applies Amount, and applies the active mask. The result keeps the source image's metadata. A derived component's reconstruction saturates an integer image at its type's range and keeps a floating image's values outside [0, 1].

Noise estimate, transform, and reconstruction

For input f, the vendored std_noise command forms a discrete Laplacian L(f), subtracts its median, and estimates

\[ \hat\sigma = \frac{1.4826\;\operatorname{median}(|L(f)- \operatorname{median}(L(f))|)}{\sqrt{20}} \]

for a 2D image. The factor is a robust Gaussian-noise scale estimate for this Laplacian stencil. One scalar estimate is used for the whole image supplied to the command; RGB therefore shares it across channels.

Let H_s be the s-level separable Haar analysis, H_s^{-1} its inverse, and c=H_s(f) a coefficient. With UI Threshold k, G'MIC uses T=k sigma and the soft shrinkage

\[ S_T(c)=\operatorname{sign}(c)\max(|c|-T,0). \]

Unlike per-layer wavelet tools, this command thresholds the complete transformed image; no separate strength exists for each scale or for the coarsest residual. Each Haar level halves the current spatial approximation and produces horizontal, vertical, and diagonal detail coefficients. Inverse Haar synthesis reconstructs the original padded grid before dimensions are restored.

Cycle spinning and blend

For N=Iterations, cycle j draws a random periodic shift D_j. The full filtered result is

\[ d=\frac{1}{N}\sum_{j=1}^{N}D_j^{-1}H_s^{-1} \left(S_T(H_s(D_j f))\right). \]

Shifting before a decimated Haar transform changes which samples meet at block boundaries. Undoing and averaging the shifts reduces translation/block artifacts. The generator is seeded with the same fixed value on every call, so the same input and settings give the same result on every run. There is no iterative convergence or tolerance.

After channel reconstruction, Amount a produces

\[ b=(1-a)f+a d,\qquad 0\le a\le1. \]

An active mask then blends f and b; mask inversion reverses its weights.

Boundary and failure behavior

• The actual scale count is min(requested, floor(log2(min(width,height))-1), 32). Very small images may not support the UI's minimum two levels and can fail in the G'MIC command.
• Dimension extension uses a zero boundary and no interpolation. High scales or non-multiple dimensions can make outer pixels differ from the interior and can increase memory substantially. Cycle shifts themselves use periodic boundaries.
• Every G'MIC result is clamped to normalized range before conversion. Integer output is rounded to the source type; double input has only float working precision. An image holding NaN/Inf is refused before the command runs.
• A G'MIC command error is reported with G'MIC's message. The source is not committed on failure.
• G'MIC receives a live cancellation flag, so Cancel can abort inside the command. Sequence cancellation also removes the temporary output before target replace.

Parameter interactions and starting points

Combination	Conservative starting point	Too weak looks like	Too strong looks like	Adjustment order
Threshold + Scales	Threshold 1.0..1.4, Scales 4..6 on a representative area	Fine grain or broader mottling remains	Stars/filaments soften, background becomes blocky, broad levels move	Choose needed scale support first, then raise Threshold slowly
Iterations + preview repeatability	10; use fewer only for faster coarse tuning	More shift-dependent/block texture	No extra “convergence”; only cost rises	Tune Threshold/Scales with moderate cycles, then increase cycles if shift artifacts matter
Amount + strong threshold	Keep Amount 1 while diagnosing, then back off toward 0.5..0.8	Denoise is visually diluted	Full result appears waxy or loses weak structure	Tune filter at full Amount, then restore source with Amount
RGB + unequal channel noise	Compare RGB with a deliberate luminance/chroma component on a duplicate	Noisiest channel remains	Cleaner channels lose detail or reconstructed color shifts (and, in an integer image, clips)	Inspect channels, choose component, then retune Threshold
Scales + image dimensions	Stay below scales that cause large padding jumps	Broad-scale noise remains	Edge/padding artifacts and memory/time jump	Reduce Scales or use a representative crop
Mask + Amount	Validate global result, then enable a feathered mask	Target area remains noisy	Mask boundary or over-smoothed protected feature becomes visible	Tune unmasked, reduce Amount, then refine mask

These are conservative source-derived starting ranges, not measured quality recommendations for a particular camera or dataset.

Inputs, outputs, and metadata

• A still produces one modified still with original width, height, channel count, and pixel type unless a converted component reconstructs RGB. G'MIC works in float and clamps to [0,1]; double precision is not retained internally.
• The result keeps the source image's header cards, astrometric solution, and other metadata for every Channel choice. No processing-state flag is changed, so a linear or nonlinear label is inherited rather than newly characterized.
• Image History records Wavelet Denoise, Threshold, Scales, Iterations, Amount, and, for a colour image, in channel: <channel>. It does not record the random shifts, mask, preview area, display stretch, G'MIC version, clamp, or actual dimension-capped scale.
• A snapshot creates an auxiliary still and leaves the target unchanged. Apply writes no user-selected file directly.
• Sequence Apply captures one parameter snapshot for all frames, processes into a temporary sequence, honors configured CFA auto-debayer on read, and replaces the source only after success. Cancellation removes the temporary output.

Performance

The dominant cost grows with cycle count, scale count, channels, and padded pixel count. Each cycle performs a forward/inverse multiscale Haar transform plus shifts and accumulation. AfterNight runs the command as one G'MIC call per run, not split across threads, so different sequence frames are expensive even when general sequence infrastructure has parallel stages.

High Scales can round each dimension up to a large 2^(scales+1) multiple. For example, scale 10 requires 2,048-sample multiples, so memory may jump rather than grow smoothly. RGB, component conversion, Amount, masks, Preview presentation, and snapshots add full-image buffers/passes. Use Preview area and fewer cycles for coarse tuning, then restore the intended cycles for the final comparison. No production timing was measured.

Worked example

• Evidence mode: Source-and-calculator-derived; not executed on project-owned astronomical data.
• Input condition: Assume the robust estimate is sigma=0.020, Threshold 1.40, Scales 6, Iterations 10, and Amount 0.50.
• Coefficient checkpoint: T=1.40×0.020=0.028. A coefficient 0.020 becomes 0; a coefficient 0.050 becomes 0.022; a coefficient -0.050 becomes -0.022 in every cycle before reconstruction.
• Blend checkpoint: If one reconstructed pixel is d=0.36 and source is f=0.40, Amount 0.50 yields 0.38 before mask blending.
• Inspect in real use: 100% smooth background, smallest stars/filaments, high-contrast edges, all image boundaries, RGB balance, and Image History.
• Limitations: The numbers demonstrate the current shrink/blend equations only. They do not predict the robust estimate, random shifts, reconstructed pixels, artifacts, quality, or runtime for real data.

Warnings and limitations

Iterations is not convergence

Iterations is cycle-spinning count, not convergence. Preview, snapshot, and Apply run the command again with the same fixed-seed shifts, so the same input and settings give the same result.

• The UI exposes one threshold for all scales and the coarse approximation; this is not a layer-by-layer wavelet denoise control.
• G'MIC uses one robust noise estimate for the supplied image. RGB channels are transformed independently with the same absolute coefficient threshold.
• More Iterations does not inherently mean more smoothing despite the current tooltip. It averages more shifted reconstructions at approximately linear cost.
• High Scales can affect large structures, force large zero-boundary padding, and produce block/ringing/edge behavior. Haar is discontinuous and particularly sensitive to sharp steps; cycle spinning reduces but does not remove this risk.
• All output is clamped to [0,1] and processed at float precision. Out-of-range float data and double-precision details are not preserved.
• An image holding NaN or infinite samples is refused; without G'MIC in the build, the process cannot run and Preview and Apply stay disabled.
• Component filtering can distort circular hue/opponent encodings and, in an integer image, clip reconstructed RGB at its type's range (a floating image keeps values outside [0, 1]). RGB thresholding can also shift color when channel noise/structure differs.
• Denoising is destructive after Apply and cannot recover clipped/calibration information. It does not update the image's linear/nonlinear processing state.

Troubleshooting

Symptom or exact error	Verified cause/validation path	Recovery	Evidence source
Increasing Iterations mostly slows processing	It is a fixed averaging count, not a convergence or repeated-strength loop	Tune Threshold/Scales/Amount for strength; use Iterations for shift averaging	WaveletDenoiseProcess.cpp (the denoise_haar command); G'MIC cycle_spinning
Broad structure or mean level changes	Soft threshold applies to the complete Haar coefficient array, including coarse approximation	Reduce Threshold and/or Scales; restore detail with lower Amount	vendored threshold and denoise_haar commands
Block patterns or damaged edges	Decimated Haar blocks, temporary zero-boundary dimension extension, and finite cycle sampling	Reduce Scales/Threshold, add cycles, use a feathered mask, or compare another denoiser	vendored Haar/resize/shift commands
Output clips out-of-range float samples	Every G'MIC result is clamped before restoring source type	Normalize/crop range before filtering and preserve a duplicate	GmicProcessSupport.cpp
Error: the image holds a sample that is not a finite number	The image has a NaN or infinite sample, which G'MIC cannot process	Repair or crop the non-finite samples, then retry; the image was left unchanged	GmicProcessSupport.cpp
Preview and Apply are disabled; tooltip says the build carries no G'MIC library	This build was made without G'MIC	Use a build that includes G'MIC, or another denoiser	GmicProcessSupport.cpp
A G'MIC error on a tiny target	Requested/UI-minimum Haar depth can exceed the dimension-derived supported depth	Use a larger image/crop or fewer supported scales where possible	vendored denoise_haar; Haar validation
Cancelled sequence leaves original unchanged	Temporary output is removed and commit is skipped	Expected; restart Apply when ready	shared sequence executor; G'MIC adapter abort flag

Related workflow steps

Finish Image Stacking, Background Extraction, and Photometric Color Calibration or Spectrophotometric Color Calibration before linear denoise. Compare the variational model in Total Variation Denoise, the learned external model in DeepSNR, and the later patch methods in NL-Bayes or Patch PCA Denoise. Continue to Histogram Transformation or another controlled stretch, then verify Image History, Image Statistics, and critical metadata.

References and provenance

• G'MIC denoise_haar reference — upstream command contract for Haar thresholding, scale selection, and cycle spinning. AfterNight bundles G'MIC 3.5.4; the exact vendored command is authoritative for this build.
• G'MIC Haar analysis and synthesis tutorial — upstream background for the separable Haar transform.
• AfterNight and vendored G'MIC project sources — authoritative for the robust estimate, soft threshold, random shifts, dimension/boundary behavior, float/clamp bridge, channels, Amount/masks, preview, metadata, history, cancellation, and sequences.