SonarWaveletNoise

SonarWaveletNoise is the rare node where the author just tells you exactly how it works, in four steps: generate noise, scale it down 50%, scale it back up to the original size, and subtract the scaled version from the original. Scaling down and back up blurs the noise, so subtracting it is essentially sharpening it - boosting the high-frequency texture and suppressing the smooth low-frequency content. Then it does that over multiple octaves, each at a different scale, and stacks the results.

That's "actual wavelet noise" in the sense of the Wikipedia concept - an additive multi-octave construction - rather than a true filterbank decomposition. Which is exactly why it's the friendly one in the wavelet family: unlike SonarWaveletFilteredNoise and SonarScatternetFilteredNoise, this node has no pytorch_wavelets dependency. It's pure scaling and subtraction, so it just works out of the box.

The inputs that matter

• octaves (default 4) - how many scale levels to stack. Each octave works at a coarser scale than the last, so more octaves = broader frequency range covered. Negative values run the octaves in reverse, which the author says "may produce weird results" - so that's basically a feature.
• initial_amplitude (default 1.0) and persistence (default 0.5) - octave strength. The first octave uses initial_amplitude, then each subsequent octave multiplies by persistence, so 0.5 gives you 1.0, 0.5, 0.25, 0.125. Lower persistence means the coarse octaves fade out faster and the result stays high-frequency.
• octave_height_factor / octave_width_factor (both 0.5) - how much each octave's noise is scaled down before the blur-and-subtract. 0.5 is the classic half-size; the two are separate so you can sharpen only one axis.
• height_factor / width_factor (both 2.0) - how much the working size shrinks after each octave. Careful: values below 1 grow the working size, and the tooltip warns it's "very easy to reach absurd latent sizes with only a few octaves."
• octave_scale_mode (adaptive_avg_pool2d) - the blur method for the downscale; octave_rescale_mode (bilinear) - the upscale; post_octave_rescale_mode (bilinear) - how each octave's output gets stretched back to the actual latent size. These are your texture-quality knobs; area and bislerp are worth a try for softer results.
• update_blend (default 1.0) and update_blend_mode (lerp) - how original - blurred blends with the original noise. With normal wavelet noise you use 100% of the sharpened result; the blend controls are there for the "what if it's only half sharpened" experiments.
• custom_noise (optional) - supply your own base noise instead of Gaussian. Fine print: all octaves get generated at maximum scale and then cropped, which the tooltip warns "may be slow."

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

What it's actually good for

Sharpened, high-frequency noise is what you want when Gaussian feels too bland - it gives the sampler a stronger high-frequency kick at the start of sampling, which can translate into crisper detail in the final image. It's also the natural base for the pack's frequency tricks: run it through SonarQuantileFilteredNoise to tame any extremes, or schedule it in with SonarScheduledNoise so it only owns part of the sampling range. Start with the defaults, then nudge persistence up if you want more mid-frequency energy or down if you want pure grit.

Installing

Ships with ComfyUI-sonar. ComfyUI Manager → search "ComfyUI-sonar" → install, or git clone https://github.com/blepping/ComfyUI-sonar into custom_nodes/, restart. No extra packages, no model files - the rare wavelet node that doesn't need pytorch_wavelets. And the pack's usual honesty applies: noise types are "best-effort," internals have shifted between versions, so pin the git revision if a particular look matters to you.