Local Histogram Equalization

Purpose and expected result

Local Histogram Equalization remaps each pixel through a cumulative distribution function (CDF) built from that pixel's own neighborhood. Unlike CLAHE, it does not divide the image into tiles, clip histogram bins, or interpolate between tile mappings. Every pixel receives a separately computed local mapping.

The expected result is a same-size still or replacement sequence with stronger local tonal separation. The shipped process always uses G'MIC's hidden Soft mode and, for default RGB, operates on Lab lightness. It can reveal faint structure, but it can also amplify noise, create local halos, and produce an overprocessed soft-light appearance.

At a glance

Item	Current behavior
Menu path	Process > Intensity Transformations > Local Histogram Equalization
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 grayscale or RGB still or sequence; optional active mask with matching geometry; a CFA still remains one grayscale mosaic plane, while a CFA sequence is auto-debayered to RGB first when its view has sequence auto-debayer enabled
Outputs	Destructively replaces the still or, after successful temporary processing, the sequence; Take snapshot creates an auxiliary still; no file is written directly
Recommended image state	Finite normalized nonlinear working copy after noise reduction; local CDF remapping is normally a post-stretch contrast operation
Preview support	Asynchronous movable-area preview, original/result switching, snapshot, display-channel selection, and separate source/result histogram; no execution autostretch
Sequence support	Yes; numeric/checkbox parameters and process Channel are captured once when Apply starts, CFA frames can be auto-debayered first, and temporary output replaces the sequence only after success
Mask support	Yes; active mask blends the Amount-mixed result with the original after local equalization
External dependencies	Bundled G'MIC 3.5.4/cgmic standard library, run through AfterNight's optional G'MIC adapter; no separate installation, model, file, or network access. 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
Processing state	Apply records the result as nonlinear under the operation local-histogram-equalization; a preview records nothing

Before you start

• Denoise before local equalization. A local CDF expands small neighborhood differences whether they are structure, shot noise, compression, or defects.
• Debayer CFA stills before color work. For CFA sequences, confirm the view's sequence auto-debayer setting: enabled frames enter as RGB; disabled frames remain a grayscale mosaic that can exaggerate the sensor pattern.
• Duplicate the image or confirm Image History is available. Apply replaces the target only after the full operation succeeds.
• Use finite normalized data. G'MIC works in float32 and the result is clamped to [0,1]; out-of-range values and double precision are not preserved. 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 default RGB if Lab-lightness enhancement is intended. Selecting a different process Channel first changes the component supplied to G'MIC.
• Preview smooth background, a bright star edge, and faint structure at 100%. Disable Preview area before Apply to check broad context and frame borders.

Procedure

1. Open Process > Intensity Transformations > Local Histogram Equalization and confirm Target Image and Channel.
2. Start at Strength 75%, Radius 4, Sigma 100, Regularization 8, Reduce Halos selected, and Amount 1.
3. Choose Radius for neighborhood scale. Increase only while the desired structure benefits more than stars, gradients, and smooth background.
4. Tune Sigma with Reduce Halos selected. Lower values restrict the CDF toward center-like intensities; higher values admit a broader intensity range.
5. Adjust Regularization to smooth an unstable or mottled equalization map. Recheck fine structure because strong bilateral smoothing can suppress it.
6. Reduce Strength to weaken the internal Soft blend; reduce Amount last to crossfade the entire completed filter result toward the source.
7. Toggle Switch image and inspect noise, halos, clipping, and color at Zoom 100%. Confirm mask/inversion, then Apply.

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
Strength (%)	Double value plus slider, percent opacity of fixed Soft blend	75.00	None	0.00..100.00; spin default step 1; slider maps 0..1000 to tenths	Raises opacity of G'MIC Soft blend between source and locally equalized layer; 0% is source, 100% is full Soft blend—not the raw CDF layer	Hidden Mode is always Soft (2); captured for execution	High values can look crunchy/flattened and amplify noise; does not avoid CDF cost at 0	Preview/Apply; selected process Channel; before final Amount/mask; still/sequence	Captured; Reset/success restore 75
Radius	Integer value plus slider, pixels	4	None	1..16; step 1; square neighborhood is (2r+1)×(2r+1)	Enlarges the local sample window from 3×3 to 33×33	Captured for execution	Larger windows include broader structures and sharply increase per-pixel work; can create broader halos	Preview/Apply; selected Channel; before blending/mask	Captured; Reset/success restore 4
Sigma	Double value plus slider, 512-bin intensity-distance scale	100.00	None	0.00..256.00; spin step 0.05 adaptive; slider maps 0..2560 to tenths; script uses 0.1+Sigma	With Reduce Halos on, higher Sigma admits more intensity differences into the local histogram; lower Sigma favors center-like levels. With it off, the complementary weighting reverses that relationship	Interacts fundamentally with Reduce Halos; captured	Very low selected Sigma can reduce support to nearly identical bins; very high values approach broad weighting. Off mode can ignore same-level samples and emphasize dissimilar neighbors	Preview/Apply; local CDF; still/sequence	Captured; Reset/success restore 100
Regularization	Double value plus slider, G'MIC bilateral spatial/range scale	8.00	None	0.00..32.00; spin step 0.05 adaptive; slider maps 0..320 to tenths; 0 disables	If nonzero, smooths the equalized result using source norm as guide, spatial sigma Regularization, and range sigma 2+Regularization in G'MIC's working convention	Applied after CDF normalization and before Soft blend	Higher values can reduce mottling/halos but blur the mapping and cost a bilateral pass; 0 preserves raw local-map variation	Preview/Apply; selected Channel; still/sequence	Captured; Reset/success restore 8
Reduce Halos	Checkbox	Selected	None	Selected/unselected	Selected uses exp(-(difference/(0.1+Sigma))²) and favors intensities similar to the center. Unselected uses 1-exp(...), gives zero weight to identical values, and increasingly favors dissimilar levels	Changes meaning of Sigma; captured	Unselected can produce aggressive/unstable local mappings; selected is the conservative path but does not guarantee halo-free output	Preview/Apply; local CDF; still/sequence	Captured; Reset/success select it
Amount	Double value plus slider, final normalized blend	1.00	None	0.00..1.00; spin step 0.05 adaptive; slider maps 0..1000 to thousandths	Linearly crossfades source and complete fixed-Soft LHE result; 0 is source, 1 is full result	Independent of Strength; captured	Reduces all visible effects but does not undo internal clipping or save CDF/bilateral work	Preview/Apply; reconstructed Channel; before mask; still/sequence	Captured; Reset/success restore 1
Histogram plot	Interactive source/result histogram	Source initially, preview result after execution	None	Mouse wheel/drag and toolbar actions below	Displays distribution and clipping diagnostics only	Updated from target/preview; separate from preview display	High bin counts add analysis/rendering cost; spatial halos are not visible in a histogram	Display only	Window lifetime
Zoom in	Histogram-toolbar action	Current histogram range	None	Click	Narrows plot range	Histogram shown	Diagnostic only	Histogram display only	Window lifetime
Zoom out	Histogram-toolbar action	Current histogram range	None	Click	Widens plot range	Histogram shown	Diagnostic only	Histogram display only	Window lifetime
Zoom reset	Histogram-toolbar action	Full histogram range	None	Click	Restores full plot range	Histogram shown	Diagnostic only	Histogram display only	Window lifetime
Show grid	Checkable histogram action	Checked	None	Checked/unchecked	Shows/hides plot grid	Histogram shown	Display only	Histogram display only	Window lifetime
Show clipping info	Checkable histogram action	Checked	None	Checked/unchecked	Shows/hides endpoint clipping readouts	Histogram shown	Binned diagnostic, not spatial-quality evidence	Histogram display only	Window lifetime
Absolute scales	Checkable histogram action	Unchecked	None	Checked/unchecked	Uses shared absolute vertical scale instead of local per-channel normalization	Histogram shown	A low-population channel can become hard to see	Histogram display only	Window lifetime
Draw histogram areas	Checkable histogram action	Unchecked	None	Checked/unchecked	Fills plotted channel areas	Histogram shown	Overlap can obscure detail	Histogram display only	Window lifetime
Show histogram	Checkable histogram action	Checked	None	Checked/unchecked	Hides/clears graph and disables toolbar, or reloads source/preview	Always available while idle	Hiding saves graph work only	Display only	Window lifetime
Plot resolution	Combo box, histogram bins	8-bit (256)	None	6-bit (64), 7-bit (128), 8-bit (256), 9-bit (512), 10-bit (1024), 12-bit (4096), 14-bit (16K), 16-bit (64K), 100 levels, 1000 levels, 10000 levels	Changes diagnostic binning only	Disabled when histogram hidden	More bins cost memory/rendering; output always uses internal 512-bin local CDF regardless	Histogram display only	Window lifetime
Target Image	Image selector	Attached/active target	None	Open eligible image or sequence	Changes source, Channel list, histogram, Preview, mask, and destination	Active target required; duplicate/exclusive launch gates apply; disabled during Apply	A wrong selection before Apply replaces the wrong target	Preview/Apply source and mask	Target lifetime; Reset unchanged
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	Default RGB is internally converted by G'MIC and equalized only in Lab lightness. A selected component is extracted first and supplied as grayscale, then reconstructed	RGB required for components; disabled during Apply; the selection is captured when Apply starts	Lab-lightness default better preserves chroma than independent RGB; alternate component round trips add buffers and possible gamut clipping	Preview/Apply; active mask post-blend; still/sequence	Dialog-local; Reset/success restore RGB
Preview	Checkbox	Checked	None	Checked/unchecked	Enables/disables asynchronous full/area execution and preview pane	Process idle	Recomputes full per-pixel CDF/bilateral after changes	Preview only; mask-aware; still/current displayed frame	Window lifetime; Reset unchanged
Preview pane	Image viewer	Filtered result in Linear mode; centered 600×600 area when possible	Execution autostretch is unavailable	Pan/scroll and toolbar actions below	Shows mask-aware result without target modification	Exists only with Preview	Holds source/result/display plus G'MIC buffers	Preview only	Window lifetime
Fit To Window	Preview-toolbar action	Preview fitted initially	None	Click or Ctrl+Shift+F	Fits whole image	Preview required	Broad context; hides fine noise	Preview display only	Viewer lifetime
Zoom In	Preview-toolbar action	Current scale	None	Click; 1.25×	Enlarges around center	Preview required	Improves local inspection	Preview display only	Viewer lifetime
Zoom Out	Preview-toolbar action	Current scale	None	Click; 0.8×	Reduces view	Preview required	More context, less fine detail	Preview display only	Viewer lifetime
Zoom 100%	Preview-toolbar action	Current scale	None	Click; scale 1.0	One-pixel viewer scale	Preview required	Preferred for noise/halos	Preview display only	Viewer lifetime
View Channel	Preview-toolbar combo box	RGB for RGB; disabled Grayscale for mono	None	Stored/derived RGB, CIE, HSL, HSV, HSI, YCbCr display components	Selects display component only	Separate from process Channel	Diagnoses color changes without altering samples	Preview display only	Viewer lifetime
Switch image	Preview-toolbar action	Result shown	None	Click; Space; right-click; Ctrl+Shift+O	Toggles result/original	Preview required	Comparison only	Preview display	Preview lifetime
Preview area	Preview-toolbar checkbox	Checked with centered 600×600 area when possible	None	Checked/unchecked; move/resize	Processes area and composites it over original preview	Preview required	Major tuning speedup; can hide broad context/border behavior	Preview/snapshot; mask cropped consistently	Preview lifetime
Take snapshot	Preview-toolbar action	No snapshot	None	Click or Ctrl+Shift+K	Re-executes and opens <target> - Local Histogram Equalization - SNAPSHOT	Preview required	Adds costly execution and auxiliary still	New still; target unchanged; area/mask honored	New view per action
Apply	Button/action	Enabled	None	One action	Captures numeric/checkbox values, requests a processing lock on the current target, executes the full still/sequence, then applies Amount/mask and commits on success	Target, Channel, and process controls lock during Apply	Per-pixel CDF, optional bilateral, conversions, mask, and sequence encoding are expensive	Destructive still/sequence replacement after success; the result is recorded nonlinear	History records six parameters and, for a colour image, the selected Channel text; then Reset
Reset	Button/action	Enabled while idle	None	One action	Restores 75/4/100/8/selected/1 and RGB	Leaves Target, Preview, histogram, mask, and preview toolbar; disabled during Apply	Checked Preview schedules another expensive execution	Configuration only	Immediate; also after successful Apply
Close	Button/action	Enabled while idle	None	One action	Closes without Apply	Disabled during Apply	G'MIC cancellation may wait for next abort check	Window/execution	Settings discarded
Sequence progress / Cancel	Shared indeterminate process/frame progress and Cancel	Created for sequence Apply	None	Cooperative Cancel	Optionally auto-debayers configured CFA frames, writes a temporary sequence, and promotes it only after success	Sequence only	Captured parameters and Channel stay fixed for every frame; high temporary disk/CPU cost	Full sequence; Amount/mask per frame	Cancel removes temporary output; success records one history step and resets

How the algorithm works

Conceptual pipeline

1. The shared process path extracts a selected component when needed, checks it for non-finite samples, and sends it through AfterNight's G'MIC adapter, one G'MIC session per run, as float32 planar fractions of full scale. For default RGB, the filter's fixed channel code 16 selects Lab lightness internally.
2. The bundled G'MIC command fx_equalize_local_histograms normalizes that working channel to integer bins 0..511. At every pixel, it scans a square Radius neighborhood using Neumann boundary extension and builds a weighted 512-bin histogram.
3. It converts the histogram to a normalized CDF and maps the center sample through that CDF. It then normalizes the resulting local-map image to 0..255.
4. If Regularization is nonzero, it applies a guided bilateral smoothing pass to the local-map result.
5. The command combines source and map with G'MIC Soft light at Strength opacity. AfterNight clamps the result to [0,1] and converts it back to the source type. On Apply it then records the result as nonlinear in the processing-state cards (ANLIN false, ANOP local-histogram-equalization), keeping every other card, before applying final Amount and mask; a preview records nothing.

Local weighted CDF

Let q(p) be the center sample quantized to one of 512 bins, and let N_r(p) be its (2r+1)×(2r+1) neighborhood. For neighbor u, define intensity difference Δ=|q(u)-q(p)| and s=0.1+Sigma.

With Reduce Halos selected, the histogram weight is

\[ w(\Delta)=\exp[-(\Delta/s)^2]. \]

Without it, the script uses the complement

\[ w(\Delta)=1-\exp[-(\Delta/s)^2]. \]

The weighted histogram and CDF are

\[ h_p(k)=\sum_{u\in N_r(p)} w(|q(u)-q(p)|)\,[q(u)=k], \]

\[ C_p(k)=\frac{\sum_{j=0}^{k}h_p(j)}{\max(10^{-5},\sum_{j=0}^{511}h_p(j))}. \]

The local-map sample is E(p)=C_p(q(p)). Selected weighting favors levels near the center; unselected weighting gives identical levels zero weight and favors increasingly dissimilar levels. This is an intensity-distance Gaussian, not a spatial Gaussian. Spatial inclusion is the hard square Radius window.

Two separate strength stages

The hidden process mode is always G'MIC Soft light. Strength s_t is its opacity between the original and locally mapped layer; it is not a linear mix with the raw CDF layer. Let the resulting G'MIC output be F. AfterNight then uses a separate linear Amount a:

\[ M(p)=(1-a)I(p)+aF(p). \]

With mask weight m(p), final output is

\[ O(p)=(1-m(p))I(p)+m(p)M(p). \]

Strength 0% and Amount 0 are both identity endpoints but neither bypasses the expensive local-CDF call. Clipping/normalization inside G'MIC occurs before Amount and mask, so later blending cannot restore lost range.

Regularization, color, and precision

• For nonzero Regularization λ, the script applies bilateral smoothing with source norm as guide, spatial sigma λ, and range sigma 2+λ in the G'MIC working convention. This smooths the map rather than changing CDF support.
• Default RGB uses Lab lightness and preserves G'MIC's reconstructed chroma. Selecting Red/Green/Blue or another process component extracts that scalar component before the fixed internal lightness selection is invoked. A derived component's reconstruction saturates an integer image at its type's range and keeps a floating image's values outside [0, 1].
• The G'MIC hand-off is float32. Integers normalize before processing; double narrows; all output clamps to [0,1] and returns to the original storage type, rounded to nearest.
• Apply records the result as nonlinear under the operation local-histogram-equalization.

Parameter interactions and starting points

Combination	Conservative starting point	Too weak looks like	Too strong looks like	Adjustment order
Radius × structure	Radius 4 on 100% preview	Structure remains compressed	Broad halos/flattening and long runtime	Radius first
Sigma × Reduce Halos	Selected, Sigma 100	Very low Sigma uses only center-like levels	Very high Sigma admits broad level population; unselected can become aggressive	Keep selected; tune Sigma second
Regularization × noise	8, compare 0 only on a duplicate	Map remains mottled/noisy	Fine local separation is smoothed away	Sigma, then Regularization
Strength × Amount	Strength 75%, Amount 1	Both low obscure whether CDF settings work	High Strength creates Soft-light artifacts even if Amount later hides them	Tune Strength, lower Amount last
RGB × component	Default RGB/Lab lightness	Luminance separation is modest	Alternate chroma/hue component produces gamut/color artifacts	Choose process Channel before CDF controls
Preview area × Radius	Area for tuning, full frame before Apply	Small area misses context	Full result surprises at borders/large structures	Tune locally, then full preview

These are source-derived checkpoints, not executed visual-quality claims.

Inputs, outputs, and metadata

• Output dimensions, channel count, pixel format, and storage type match the source still/frame. Integer results quantize; float/double still clamp and double passes through float32.
• Still Apply modifies only after success. Take snapshot creates a new still; histogram/display settings are not baked into samples.
• The result keeps the source image's header cards and metadata, and Apply records its processing state as nonlinear under the operation local-histogram-equalization, with a lineage entry naming the process. A preview records nothing. Image History uses the process name.
• History records Strength, Radius, Sigma, Regularization, Amount, Reduce Halos, and, for a colour image, in channel: <channel>. It omits fixed Soft mode, internal Lab-lightness code, mask/inversion, G'MIC version, and CDF bins.
• Successful Apply resets process parameters/RGB. Target, Preview, histogram, mask, and preview toolbar remain.
• Sequence Apply promotes a temporary sequence only after success; Cancel removes temporary output. There is no direct output path/name control.
• CFA sequence frames are auto-debayered before filtering when sequence auto-debayer is enabled; otherwise the filter receives the mosaic plane.

Performance

• Per pixel, the script scans (2r+1)² samples and cumulatively scans 512 bins. A useful model is O(width × height × ((2r+1)² + 512)) per processed scalar channel, plus color conversion, normalization, and blending.
• Radius grows the neighborhood from 9 to 1089 samples, so it is the dominant user-controlled cost. The fixed CDF scan remains large even at Radius 1.
• Nonzero Regularization adds a full guided bilateral pass. Default RGB also adds Lab conversion/reconstruction.
• Memory includes G'MIC float images, local command/list intermediates, destination, color conversion, mask, preview, histogram, and temporary sequence storage.
• Preview area is the safest prototype path. Strength 0, Amount 0, and unchecked Preview do not provide an algorithmic fast path for Apply.
• Focused tests illustrate cost even on synthetic small/medium fixtures: the four-test filter took 57.166 seconds in Release and 105.196 seconds in Debug on the recorded machine. These are validation timings, not production benchmarks.
• No approved astronomical timing or peak-memory measurement was made.

Worked example

• Evidence mode: Source-and-equation-derived; not executed on approved astronomical data.
• Input/settings: Finite normalized nonlinear image, default RGB/Lab lightness, Strength 75%, Radius 4, Sigma 100, Regularization 8, Reduce Halos selected, Amount 1.
• CDF checkpoint: If a local weighted histogram has total weight 10 and cumulative weight 7 at the center bin, its local-map value before global normalization/blending is 0.7.
• Endpoint checkpoint: Strength 0% returns source at the G'MIC blend stage; Amount 0 returns source at the AfterNight stage.
• Inspect: At 100%, check smooth background noise, star rims, mottling, border behavior, hue/chroma stability, clipping bins, and recorded history.
• Limitations: No visual improvement, executed sequence output, production timing, or memory result is claimed.

Warnings and limitations

• Local Histogram Equalization is not CLAHE: no tiles, clip limit, or tile interpolation exist. It avoids tile seams by computing a CDF at every pixel, at much higher cost.
• The only process mode is hidden Soft. Raw and Hard remain disabled because the process source records them as broken for 16-bit images.
• Sigma measures intensity difference in 512-bin space, not spatial distance. Radius alone defines the square spatial neighborhood.
• Turning Reduce Halos off does more than remove protection: it complements the weighting, gives identical values zero weight, and favors dissimilar levels.
• Default RGB operates on Lab lightness, not independently on R/G/B. Alternate process components can alter color/gamut during reconstruction.
• Local equalization can amplify noise, ringing, hot pixels, compression, and gradient residuals. Regularization and final blending do not make bad input safe.
• Every result is clamped to [0,1]; out-of-range float data and double precision are not preserved. An image holding NaN or infinity is refused by name; without G'MIC in the build, the process cannot run and Preview and Apply stay disabled.
• History uses the process name but omits the hidden mode/internal channel details.

Troubleshooting

Symptom or exact error	Verified cause/validation path	Recovery	Evidence source
Result looks unlike CLAHE	This process computes per-pixel weighted CDFs and Soft blend; no tiles/clip limit	Use CLAHE when tiled contrast-limited behavior is required	Bundled G'MIC command
Smooth background becomes mottled	Neighborhood CDF expands small differences/noise	Denoise first; reduce Strength/Radius; increase Regularization cautiously	CDF/cost model
Halos appear with Reduce Halos selected	Similarity weighting reduces but does not eliminate local CDF and Soft-light edge effects	Reduce Strength/Radius; inspect full frame; use mask	Weighted CDF and Soft blend
Turning Reduce Halos off changes result drastically	Weight becomes 1-exp(...), not unweighted histogram	Re-enable it; retune Sigma only if the complementary behavior is intentional	Exact script equation
Sigma seems opposite with protection off	Complementary weighting favors larger intensity differences	Re-enable Reduce Halos or reason from the off-mode equation	Exact script equation
Colors shift	Non-default component reconstruction or extreme Lab-lightness mapping clips/gamut shifts	Return process Channel to RGB; reduce Strength/Amount	Channel pipeline
Process is very slow	Per-pixel window plus 512-bin CDF, optional bilateral, and preview repeats	Use small Preview area, lower Radius, pause changes	Complexity model
Preview differs from Apply	Preview area, Channel, mask, or display conditions changed	Match controls, disable area	Shared execution paths
The image is marked nonlinear after Apply	Apply records local equalization as a nonlinear operation	Expected; a later linear-only process will see the recorded state	Processing-state record
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 CLAHE	GmicProcessSupport.cpp
A G'MIC error is reported	G'MIC failure, invalid data, or memory exhaustion	Verify finite normalized data; retry smaller duplicate	Process error path

Related workflow steps

Complete noise reduction before local equalization. Compare Local Contrast Enhancement for a high-radius residual model, and CLAHE for tiled CDFs with a clip limit and interpolation. Use Masks to protect smooth background and Image History to compare or revert Apply.

References and provenance

• G'MIC blend reference — official Soft light and opacity command contract; accessed 2026-07-14.
• G'MIC bilateral reference — official guide/spatial/range sigma contract; accessed 2026-07-14.
• G'MIC apply_channels reference — official channel-selection and 0–255 working convention; accessed 2026-07-14.
• David Tschumperlé's 2018 G'MIC implementation note — primary author statement that this local-histogram filter is not CLAHE and source listing for the original controls.
• Bundled G'MIC 3.5.4 gmic_stdlib.gmic and AfterNight sources/tests are authoritative for the shipped CDF, fixed mode/internal channel, controls, G'MIC hand-off, masks, state/history, preview, and sequences.