SonarWaveletFilteredNoise

Wavelet decomposition splits a signal into a low-frequency part (the smooth, coarse structure) and a high-frequency part (the fine detail). SonarWaveletFilteredNoise applies that split to your noise and lets you treat the two halves differently - damp the low end, boost the high end, or even use entirely different noise sources for each side. It's a frequency-domain equalizer for the chaos your sampler starts from.

Where the pack's simpler SonarWaveletNoise just sharpens noise by subtracting a blurred copy, this node does a real wavelet decomposition and gives you control over the bands. That's the difference between a one-knob effect and an actual mixing desk.

Requirements first

This node needs the pytorch_wavelets package in the same Python environment ComfyUI runs in - same as SonarWaveletCFG and SonarScatternetFilteredNoise. If ComfyUI is in a venv or portable build, install there:

/path/to/your/python -m pip install pytorch_wavelets

Without it, the node errors when it tries to run.

The inputs that matter

• custom_noise (optional) - the source noise. Unconnected defaults to Gaussian.
• custom_noise_high (optional) - the interesting one. If you connect it, this noise is used for the high-frequency side of the wavelet while custom_noise provides the low-frequency side. That means you can literally run a smooth, low-frequency noise type on the bottom end and a wild, high-frequency type on top - a way to build hybrid noise that no single generator produces. Unconnected, the high side just reuses the same source.
• normalize_noise - normalize the source before filtering.
• normalize - the standard default/forced/disabled rebalance control.
• yaml_parameters - a text input for the wavelet config (wavelet type, level, padding mode, how the bands get scaled and blended). There's no hand-holding here: the tooltip notes "there is generally not much error checking." If you're not ready for YAML, leave it empty and use the defaults.

Output is a SONAR_CUSTOM_NOISE feeding a Sonar sampler, SamplerConfigOverride, or NoisyLatentLike.

Where it fits

This is the practical middle ground of the pack's wavelet family. SonarScatternetFilteredNoise is more powerful but slow, experimental, and image-only; SonarWaveletNoise is a blunt sharpening tool. This node is the one to reach for when you want frequency control without the scatternet's complexity - especially the dual-noise trick, which is genuinely its own thing and hard to replicate any other way. The wavelet lineage here traces back to the same ideas RES4LYF popularized for sampler-side detail work, so if you already like that school of "shape the noise, shape the image," you'll feel at home.

Installing

It ships with ComfyUI-sonar. Install the pack via ComfyUI Manager (search "ComfyUI-sonar") or git clone https://github.com/blepping/ComfyUI-sonar into custom_nodes/, restart, then add pytorch_wavelets as above. No model files.