SonarQuantileFilteredNoise

Clip the extremes out of your noise

By blepping·Created 3 years ago·Updated 15 days ago· 49

custom_noise

SONAR_CUSTOM_NOISE

◄factor1.000►

◄quantile0.850►

◄dim1►

◄flattentrue►

◄norm_factor1.000►

◄norm_power0.500►

◄normalize_noisefalse►

◄normalizedisabled►

◄strategyclamp►

Gaussian noise has a long tail. Most of the values sit near zero, but a few extreme outliers show up in every batch, and when they land they act like little bombs in your latent - over-bright spots, speckle, banding that wasn't there in the prompt. SonarQuantileFilteredNoise is a custom noise filter that measures the noise's own distribution and trims those outliers, so what reaches the sampler is a tamer, better-behaved version of whatever noise you fed it.

It works the way the name suggests: compute the quantile bounds of the noise (0.85 by default means the top 15% of values count as extreme), then treat everything beyond the bounds however you like. The default quantile of 0.85 is the author's recommendation for a reasonable starting point - "it really depends on the input and how many of the values are extreme," which is his polite way of saying tune it and look.

The inputs that matter

custom_noise - what to filter. Feed it a SonarCustomNoise chain or another custom-noise node. This is the source you're taming.
quantile (default 0.85) - where the cutoff sits. 1.0 or 0.0 disables quantile normalization entirely. A negative value flips the logic and treats values closest to zero as the "extremes" - explicitly experimental, but it's how you build noise that's mostly zeros with occasional spikes.
strategy - the big one, with 43 choices. clamp is the sane default: outliers get pinned to the quantile bounds. The rest are experimental reshaping modes - tanh, sigmoid, sin and friends squash outliers instead of pinning them, and zero / reverse_zero are for when you plan to add this noise to something else (zero zeroes everything outside the range; reverse_zero keeps only the outliers). Unless you have a specific effect in mind, leave it on clamp.
dim - what dimensions the quantile normalization uses. "global" flattens everything; otherwise it's a dimension index starting at 0. For image latents that's batch=0, channel=1, row=2, column=3. Video latents add a frame dimension.
flatten (default true) - flattens before normalizing. Turn it off and rows/columns get a strong influence on the result; the author warns this can look weird.
norm_factor / norm_power - applied right around the clipping step. norm_factor scales the input just before clipping, norm_power raises the absolute value to a power after. Both "generally should be left at the default."
factor - overall strength of the generated noise; normalize - the standard default/forced/disabled rebalance-to-1.0 control.

Where it slots in

The output is a SONAR_CUSTOM_NOISE, so it behaves like any other link in the pack's noise chain: feed it into a Sonar sampler, a SamplerConfigOverride with a custom noise input, or NoisyLatentLike. A popular pattern is putting it after an extreme noise type - take a Pyramid or Voronoi source that's too wild, run it through quantile filtering, and get the character without the blowouts.

Installing

It's part of ComfyUI-sonar - install the pack once via ComfyUI Manager (search "ComfyUI-sonar") or git clone https://github.com/blepping/ComfyUI-sonar into custom_nodes/, restart, done. No extra Python packages needed for this one.

One honest warning from the pack's own changelog: quantile normalization internals have been reworked a few times, and the author notes those fixes "will likely change seeds." So if you're chasing a specific look across updates, pin the git revision. Also, if your outputs suddenly look different after updating the pack, the quantile math is the first suspect.

Categoryadvanced/noise

Inputs (10)

Name	Type	Default	Description
factor	FLOAT	1.000-10000–10000	Scaling factor for the generated noise of this type.
custom_noise	SONAR_CUSTOM_NOISE,OCS_NOISE		Custom noise type to filter. The following input types are supported: SONAR_CUSTOM_NOISE, OCS_NOISE
quantile	FLOAT	0.850-1–1	When enabled, will normalize generated noise to this quantile (i.e. 0.75 means outliers >75% will be clipped). Set to 1.0 or 0.0 to disable quantile normalization. A value like 0.75 or 0.85 should be reasonable, it really depends on the input and how many of the values are extreme. (Experimental) You can also use a negative quantile to consider values closest to 0 to be 'extreme'.
dim	COMBO	1	Controls what dimensions quantile normalization uses. Dimensions start from 0. Image latents have dimensions: batch, channel, row, column. Video latents have dimensions: batch, channel, frame, row, column.
flatten	BOOLEAN	true	Controls whether the noise is flattened before quantile normalization. You can try disabling it but they may have a very strong row/column influence.
norm_factor	FLOAT	1.0000.00001–10000	Multiplier on the input noise just before it is clipped to the quantile min/max. Generally should be left at the default.
norm_power	FLOAT	0.500-10000–10000	The absolute value of the noise is raised to this power after it is clipped to the quantile min/max. You can use negative values here, but anything below -0.3 will probably produce pretty strange effects. Generally should be left at the default.
normalize_noise	BOOLEAN	false	Controls whether the noise source is normalized before quantile filtering occurs.
normalize	COMBO	disabled	Controls whether the generated noise is normalized to 1.0 strength after quantile filtering.
strategy	COMBO	clamp	Determines how to treat outliers. zero and reverse_zero modes are only useful if you're going to do something like add the result to some other noise. zero will return zero for anything outside the quantile range, reverse_zero only _keeps_ the outliers and zeros everything else.

Outputs (1)

Name	Type	Description
SONAR_CUSTOM_NOISE	SONAR_CUSTOM_NOISE	A custom noise chain.