Noise Generator

Purpose and expected result

Noise Generator adds synthetic random noise to a working image so developers and advanced users can test denoisers, masks, rejection behavior, and other noise-sensitive processing. It is a test-data tool, not a simulation of a full camera noise model: it does not combine calibrated read noise, shot noise, dark current, gain, and quantization from acquisition metadata.

The result is a same-size replacement image or sequence with one of seven noise selections - six distinct noise models drawn by AfterNight's own engine, plus LocalVar, which is Gaussian noise of variance 0.01 - applied to RGB, one stored channel, or one converted component. The dialog draws a random seed when it opens and holds it for every preview, snapshot, and Apply, so the preview shows the noise Apply adds. After an Apply completes, the dialog draws a new seed, so a second Apply adds new noise. There is no seed control, and the seed is not recorded.

At a glance

Item	Current behavior
Menu path	Process > Utilities > Noise Generator
Enabled when	An image or sequence is active, another Noise Generator is not already attached to it, and no attached process holds exclusive access
Inputs	One grayscale or RGB still/sequence and optional matching active mask; configured CFA sequence frames can be auto-debayered before processing
Outputs	Replaces the still or, after successful temporary processing, the sequence; Take snapshot creates an auxiliary still; no file or seed record is written
Recommended image state	A duplicate or synthetic finite normalized test image; linear data is easiest to interpret, but the process accepts the current stored state
Preview support	Asynchronous full-image or movable-area preview, original/result switching, snapshot, display channels, and display-only autostretch; every recomputation draws from the dialog's held seed, so the same settings show the same noise
Sequence support	Yes; noise mode, numeric values, Channel, and seed are captured once when Apply starts; each frame draws its own noise from the seed shifted by its frame index; temporary output replaces the sequence only after success
Mask support	Yes; the active mask blends the generated result with the original after noise generation
External dependencies	None; the AfterNight engine's own noise generator. No Python interpreter, scikit-image, global lock, or network access

Before you start

• Duplicate valuable data. Successful Apply replaces the target, and the seed is not recorded, so the generated noise cannot be subtracted later.
• Decide what is being tested. Gaussian and Speckle expose Mean/Var; impulse modes expose Amount; Poisson is data-dependent; LocalVar is Gaussian noise at a fixed variance of 0.01.
• Use normalized finite data. Mean, Var, and Amount are fractions of full scale: an integer image's values are restated in its own codes (Mean times the type's span, Var times its square). Generated output is clipped to the image's full scale - [0,1] for a float image, the type's range for an integer one - and a pepper sample takes the lowest value of that range.
• For RGB, decide between RGB and a component. RGB generates independent random values for R, G, and B samples; impulse noise therefore need not replace all three samples at the same pixel.
• The full-frame Preview shows the noise the next Apply adds: the dialog holds one seed until an Apply completes. Changing a setting changes the noise computed from that seed; reopening the dialog or completing an Apply draws a new seed.

Procedure

1. Activate a duplicate or synthetic target and open Process > Utilities > Noise Generator. Confirm Target Image, Channel, and mask state.
2. Keep Preview checked and select Noise Type. Use the enabled numeric controls listed below; controls that remain enabled after a mode switch can be ignored when the backend does not consume them.
3. For Gaussian/Speckle, choose Mean and Var. For Salt/Pepper/Salt & Pepper, choose Amount and, for the combined mode, Salt vs Pepper. Poisson and LocalVar have no effective numeric control in this UI.
4. Inspect the preview at Zoom 100%. Confirm the range and spatial character. Use Switch image to compare the current draw with the original.
5. Uncheck Preview area for a full-frame distribution check. If using a mask, confirm its enabled/inverted state and geometry.
6. Select Apply. The full-frame result is the noise the preview showed; the dialog then draws a new seed for the next run. 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
Noise Type	Combo box	Gaussian	None	Gaussian, LocalVar, Poisson, Salt, Pepper, Salt & Pepper, Speckle	Selects the distribution and which numeric arguments are consumed	Disabled during Apply; mode changes update numeric enablement, but Salt/Pepper/Salt & Pepper do not explicitly disable Mean/Var and Speckle does not explicitly disable Salt vs Pepper, so stale enabled controls can remain after some transitions	Every mode draws per sample from the held seed; changing type recomputes the preview	Preview/Apply; selected channel/component; before mask; still/sequence	Captured per execution; Reset and successful Apply restore Gaussian
Mean	Double value plus slider, additive-noise mean as a fraction of full scale (Speckle: relative multiplier mean)	0.0000	None	0.0000..0.5000; configured single-step hint 0.005, but adaptive decimal stepping makes arrow/keyboard increments vary with magnitude; slider has 1001 positions	Raises the Gaussian additive offset or Speckle multiplier mean	Intended for Gaussian and Speckle; ignored by every other backend mode even if stale-enabled after a mode transition	Positive means brighten Gaussian output and bias Speckle in proportion to source; clipping increases near the upper bound	Preview/Apply; selected channel/component; before mask; still/sequence	Captured per execution; Reset/successful Apply return 0
Var	Double value plus slider, variance	0.0020	None	0.0000..0.5000; configured single-step hint 0.001, but adaptive decimal stepping makes arrow/keyboard increments vary with magnitude; standard deviation is sqrt(Var)	Widens Gaussian/Speckle random variation	Intended for Gaussian and Speckle; ignored by LocalVar and every other backend mode even if stale-enabled	Larger variance increases clipping and obscures structure; 0 removes random spread but Mean can still change samples	Preview/Apply; selected channel/component; before mask; still/sequence	Captured per execution; Reset/successful Apply return 0.002
Amount	Double value plus slider, replacement probability per selected array sample	0.0200	None	0.0000..0.5000; configured single-step hint 0.001, but adaptive decimal stepping makes arrow/keyboard increments vary with magnitude	Raises the probability that Salt/Pepper/Salt & Pepper replaces a sample	Enabled for the three impulse modes; ignored by Gaussian, LocalVar, Poisson, and Speckle	More replacements destroy more source samples; RGB draws independently per channel sample	Preview/Apply; selected channel/component; before mask; still/sequence	Captured per execution; Reset/successful Apply return 0.02
Salt vs Pepper	Double value plus slider, salt fraction among selected impulse replacements	0.50	None	0.00..1.00; configured single-step hint 0.01, but adaptive decimal stepping makes arrow/keyboard increments vary with magnitude; 0 all pepper, 1 all salt	Raises the share replaced by the top of the range; lowers the share replaced by its bottom	Consumed only by Salt & Pepper; may remain stale-enabled when switching from Salt & Pepper to Speckle, where it is ignored	Extreme values make one-sided impulse tests; does not change total Amount	Preview/Apply; selected channel/component; before mask; still/sequence	Captured per execution; Reset/successful Apply return 0.5
Target Image	Image selector	Attached/active target	None	Open eligible still or sequence	Changes source, Channel list, Preview, active mask, and Apply destination	Active target required; duplicate/exclusive launch gates apply	A wrong target is destructively replaced after success	Preview/Apply source and active mask	Target lifetime; Reset does not change it
Channel	Combo box	RGB for RGB; disabled Grayscale for mono	None	RGB, Red, Green, Blue, CIE Lab/Lch/XYZ, HSL, HSV, and YCbCr components exposed by the shared selector	RGB draws noise independently for every stored channel sample; a component draws on one converted plane and reconstructs RGB	Components require RGB; disabled during Apply; the selection is captured when Apply starts	Component conversion adds buffers/quantization/gamut risk	Preview/Apply; active mask post-blend; still/sequence	Dialog-local; Reset/successful Apply restore RGB on RGB targets
Preview	Checkbox	Checked	None	Checked/unchecked	Enables/disables asynchronous full-image or selected-area random generation	Process idle	Every recomputation draws from the dialog's held seed	Preview only; mask-aware; still/current displayed sequence frame	Window lifetime; Reset unchanged
Preview pane	Image viewer	Current generated result	None	Pan/drag and toolbar actions below	Shows the current mask-aware realization without modifying target	Exists only with Preview	Holds source/result/display buffers	Preview only	Window lifetime
Fit To Window	Preview-toolbar action	Fitted initially	None	Click or Ctrl+Shift+F	Fits the whole image	Preview image required	Useful for broad distribution; resampling hides individual impulses	Preview display only	Viewer lifetime
Zoom In	Preview-toolbar action	Current scale	None	Click; 1.25× relative zoom	Enlarges around viewport center	Preview image required	Improves sample inspection; increases panning	Preview display only	Viewer lifetime
Zoom Out	Preview-toolbar action	Current scale	None	Click; 0.8× relative zoom	Reduces around viewport center	Preview image required	Adds context but hides fine grain	Preview display only	Viewer lifetime
Zoom 100%	Preview-toolbar action	Current scale	None	Click; absolute scale 1.0	Shows one-pixel display scale	Preview image required	Preferred for impulse locations and channel-wise grain	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, and YCbCr display components	Selects only displayed component	Separate from process Channel; components require RGB	Helps reveal independent per-channel draws without changing data	Preview display only	Viewer lifetime
Display Mode	Preview-toolbar combo box	Inherits target Linear/Autostretch	None	Linear, Autostretch	Changes display only; Autostretch takes the image's exact median and MAD and rounds each display byte to the nearest value	Autostretch enables settings action	Stretch can exaggerate low-amplitude noise	Preview display only	Preview-view state
Autostretch Settings	Preview-toolbar dialog action	Disabled in Linear mode; inherits target-view settings	Shared-view values load when opened	Opens preview autostretch controls; Apply Stretch hidden	Tunes display only	Autostretch mode only	Changes judgment, never generated samples	Preview display only	Viewer-scoped
Switch image	Preview-toolbar action	Generated result shown	None	Click; Space; right-click; Ctrl+Shift+O	Toggles current result/original	Preview required	Comparison only; no recomputation	Preview display	Preview lifetime
Preview area	Preview-toolbar checkbox	Checked with initial 1200×1200 area	None	Checked/unchecked; move/resize overlay	Generates noise only for the selected area and composites over original preview	Preview required	Reduces tuning cost; the area is drawn on its own, so judge the distribution there and the exact samples in a full-frame preview	Preview/snapshot; mask cropped consistently; Apply always full-frame	Preview lifetime
Take snapshot	Preview-toolbar action	No snapshot	None	Click or Ctrl+Shift+K	Re-executes with the dialog's held seed and opens <target> - Noise Generator - SNAPSHOT	Preview required	Adds a complete execution and auxiliary still; it draws the same noise as the preview of the same scope	New still; target unchanged; area/mask honored	Each action creates a view; does not draw a new seed
Apply	Button/action	Enabled	None	One action	Captures noise type, numeric values, Channel, and seed, locks the target, and draws the full target from the held seed	Target must lock; process controls are disabled during Apply; another Apply is ignored while running	Component conversion, masks, sequence encoding, and temporary storage add cost	Destructive still/sequence replacement after success	History records Noise Type and, for a colour image, the Channel label, not numeric values or seed; settings then Reset and a new seed is drawn
Reset	Button/action	Enabled while idle	None	One action	Restores Gaussian, Mean 0, Var 0.002, Amount 0.02, Salt vs Pepper 0.5, and RGB for RGB targets	Does not change Target, Preview, mask, preview toolbar, or the held seed; Preview checked recomputes	Cannot restore a seed an earlier Apply used	Configuration only	Immediate; also invoked after successful Apply
Close	Button/action	Enabled while idle	None	One action	Closes without applying current settings	Disabled during Apply	Requests preview cancellation, which the generator observes	Window/execution	Settings and the held seed are discarded
Sequence progress / Cancel	Shared per-process/frame progress and Cancel	Created for sequence Apply	None	Cooperative Cancel; the generator observes it while drawing a frame	Sequence target only	A frame stops promptly on Cancel	Each frame draws its own noise from the seed shifted by its frame index; captured mode, numerics, and Channel are stable	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 handed to AfterNight's engine in its stored units. Mean and Var are restated in those units for an integer type; the clip range and the salt and pepper values are the type's full scale ([0,1] for float).
2. The engine draws one random value or replacement decision per sample. Every draw is keyed on the sample's coordinate and the run's seed rather than taken from a generator that advances, so the same seed gives the same noise at any thread count. A sequence frame uses the seed shifted by a fixed multiple of its frame index.
3. The result is clipped to the full-scale range and written back in the source pixel type.
4. AfterNight reconstructs RGB for a component path, keeps the source metadata, then applies the active mask blend.

Model and equations

Let I(p) be a selected sample as a fraction of full scale, m be Mean, v be Var, a be Amount, q be Salt vs Pepper, and N(m,v) be an independent Gaussian draw with mean m and variance v.

• Gaussian: O(p) = I(p) + N(m,v).
• LocalVar: O(p) = I(p) + N(0,0.01) at every element - the Gaussian model at a fixed variance of 0.01. Mean/Var are disabled/ignored for this mode.
• Speckle: O(p) = I(p) + I(p)N(m,v).
• Salt: each element is replaced by the top of the range (1) with probability a.
• Pepper: each element is replaced by the bottom of the range (0) with probability a.
• Salt & Pepper: replacement occurs with probability a; a replaced element becomes the top with conditional probability q, otherwise the bottom.
• Poisson: each sample is redrawn as Poisson(I(p)s)/s with a rate scale s fixed by the sample type: 256 for an 8-bit integer image and 65,536 for every other type, so the noise depends on the sample value and not on the rest of the picture. A sample at or below zero becomes zero.

All array elements are independent draws under these models. In an interleaved RGB array, p identifies a channel sample, not an RGB triplet. The active mask is applied later by the shared process path:

\[ F(p)=(1-M(p))I(p)+M(p)O_c(p), \]

where M is the effective normalized mask (or its inverse) and O_c is the clipped, type-restored generated result.

Reproducibility, boundaries, and failure behavior

• The dialog draws its seed from the system's entropy source and the clock when it opens, holds it for every preview, snapshot, and Apply, and draws a new one after an Apply completes. Equal settings, input, and seed give the same noise at any thread count; the seed is not stored with the image, so a later session cannot reproduce it.
• Var is variance, not standard deviation. The Gaussian sigma is sqrt(Var).
• Clipping occurs before mask blending. Out-of-range random values can be flattened irreversibly; integer conversion then rounds them.
• NaN/Inf are not validated and stay non-finite.
• Generator failures are reported as an error; the target is unchanged.
• The generator observes Cancel while it runs, and no global lock serializes it with other processes.

Parameter interactions and starting points

Combination	Conservative starting point	Too weak looks like	Too strong looks like	Adjustment order
Gaussian Mean × Var	Mean 0, Var 0.002 on a duplicate	Distribution is indistinguishable at the chosen display scale	Broad clipping and strong brightness dispersion	Keep Mean 0; raise Var; add Mean only for an intentional bias
Speckle Mean × Var	Mean 0, low Var	Bright structures show little signal-dependent variation	Bright structures fragment/clip much more than dark background	Set Mean, then raise Var while comparing bright/dark regions
Amount × Salt vs Pepper	Amount 0.02, ratio 0.5	Too few impulses to stress the tested operation	Source structure is dominated by replacements	Set total Amount, then tune the salt/pepper split
RGB × component	Compare RGB with one deliberate luminance component	One plane does not stress color-channel handling	Independent RGB impulses create colored samples	Choose channel semantics first, then distribution/strength

These are test-design starting points, not measurements of astronomical quality. Reopen the dialog, or Apply to a fresh duplicate after a completed Apply, to get further independent draws when evaluating statistical behavior.

Inputs, outputs, and metadata

• Dimensions normally match input. Channel count normally matches, except that configured CFA sequence frames can be auto-debayered to RGB before generation.
• Output keeps the source pixel type. Noise is drawn in the stored units and integer data is rounded; RGB/component reconstruction can add further rounding/gamut clipping.
• Apply replaces a still only after a successful result. Preview and snapshot do not modify it. A snapshot is the held seed's draw in a separate still view.
• The result keeps the source image's header cards and metadata. Image History records Noise Type and, for a colour image, in channel: <channel>; it does not record Mean, Var, Amount, Salt vs Pepper, or the seed.
• No explicit linear/nonlinear state or simulated-camera provenance is added.
• Sequence Apply writes temporary output and draws separately per frame from the held seed shifted by the frame index. Success replaces the sequence; cancellation removes temporary output. No seed manifest, noise map, report, or standalone file is created.

Performance

• Generation scales with selected pixel count and channel count and runs on the threads the operation is granted; RGB triples the sample work relative to mono.
• Poisson draws take a bounded number of uniform draws per sample whatever the rate, so they cost more than the other modes but cannot run unbounded.
• No global lock serializes the generator with other processes.
• Preview area reduces tuning work. No production timing or peak-memory measurement was run.

Worked example

• Evidence mode: Source-and-test-derived; not executed on approved astronomical data.
• Input condition and settings: Constant finite mono image at 0.5, Gaussian, Mean 0, Var 0.002, no mask.
• Expected checkpoint: A generated result differs from the constant source, is clipped to [0,1], and has an intended standard deviation of sqrt(0.002) ≈ 0.0447 before finite-sample variation/clipping.
• Reproducibility checkpoint: Trigger two full-frame previews, or a full-frame Preview then Apply. With the same settings the sample values match, because both draw from the dialog's held seed; a second Apply after the first completes draws different noise.
• Inspect at 100%/history: Confirm grain on the selected process Channel and verify history records type/channel only.
• Limitations: No claim is made for exact sample statistics, visual realism, production runtime, sequence output, or memory without an approved execution.

Warnings and limitations

• There is no seed control or seed history. The held seed reproduces noise only within one dialog session, until an Apply completes.
• This is not a calibrated sensor model and does not read gain, exposure, temperature, read-noise, dark-current, or camera quantization metadata.
• LocalVar is not spatially varying; it is Gaussian noise of constant variance 0.01. The visible Mean/Var values do not affect it.
• Mode switches can leave ignored controls enabled: Gaussian to Salt/Pepper/Salt & Pepper can leave Mean/Var enabled; Salt & Pepper to Speckle can leave Salt vs Pepper enabled. Backend routing, not enabled appearance, determines effect.
• RGB randomness is per channel sample. Salt/pepper can produce colored impulses.
• Output clips and returns to source type; clipping and integer rounding are irreversible.

Troubleshooting

Symptom or exact error	Verified cause/validation path	Recovery	Evidence source
Apply does not match a Preview area	The area is drawn on its own	Compare a full-frame preview with Apply	Coordinate-keyed draws
A second Apply adds different noise	The dialog draws a new seed after each completed Apply	Expected; a completed Apply's seed cannot be reused, so duplicate the noisy result when identical noise is needed elsewhere	executeProcessCompletedImpl()
Mean/Var appears enabled but has no effect	Stale control enablement after switching to an impulse mode; the generator ignores the values	Use only the mode-consumed controls in this page; Reset before changing modes if clarity is needed	updateEnabledActions() and process parameter routing
LocalVar looks like uniform-strength Gaussian noise	LocalVar is Gaussian noise of constant variance 0.01	Use Gaussian for a tunable uniform variance; use another tool for a spatial variance map	Process parameter routing
RGB salt/pepper contains colored pixels	Replacement decisions are drawn for every channel sample independently	Select a luminance/component channel if impulses should share one reconstructed plane	Per-sample draws; Channel dispatcher
Highlights/shadows flatten	Generated values are clipped to full scale before type conversion/mask blend	Reduce Mean/Var/Amount and work on a duplicate	Full-scale clip
Result is unchanged	Gaussian/Speckle Var 0 with Mean 0, impulse Amount 0, or mask protection	Check mode-consumed values and mask/inversion	Distribution equations; shared mask blend
An error is reported	Invalid data or insufficient memory	Verify target data; retry on a duplicate	Process error path

Related workflow steps

Use this process to build controlled stress inputs before comparing Wavelet Denoise, Total Variation Denoise, or stacking rejection. Use Pixel Math when a formula with a stated seed is required, and Image History to compare or revert destructive Apply. Save test fixtures through Saving and Exporting, noting that this process does not preserve its seed for later recreation.

References and provenance

• scikit-image 0.26.0 random_noise — the noise modes and parameters whose distributions the engine's generator follows; this process does not use scikit-image or Python.
• AfterNight source and focused tests — authoritative for menu/controls, exact parameter routing, fixed LocalVar behavior, the held seed, type/channel/mask/metadata/history behavior, and sequences.