Fourier Filter Node

The waveform generators in the SigmaWaveFormNodes pack are great at making jagged, spiky noise schedules. The problem is the jagged, spiky part. The Fourier Filter Node is the pack's answer: take any SIGMAS sequence, decompose it into frequencies, carve out the ones you don't want, and stitch it back together. It's the post-processing half of the pack - generate a wave with Sigma Waveform Node, then run it through this to knock the rough edges off before it ever reaches a sampler.

This is a real, mechanically sensible tool, and it's the one node in the pack you could justify using on sigmas that didn't come from this pack at all. If a custom schedule from anywhere else feels jittery - one step wildly louder than its neighbors - a lowpass here will settle it down.

How it works

Genuine Fourier filtering, not a pretend version. It takes your input sigmas, runs an FFT to get frequency-space coefficients, computes the matching frequency axis with fftfreq, and then zeroes out the bins that fall outside your filter's passband. If apply_inverse_fourier is on (it is by default), it runs the inverse FFT and hands back a real, smooth sequence.

The filter_type choices are the standard four:

• lowpass (default) - keeps the slow-moving shape, kills the high-frequency jitter. This is the one you actually want for smoothing a waveform schedule.
• highpass - the opposite: keeps only the fast wobble. Mostly useful for studying, not generating.
• bandpass / bandstop - keep or kill a slice of frequencies. The band is computed as (cutoff_frequency, cutoff_frequency * 2), so with the default cutoff of 0.1 the passband is roughly 0.1–0.2.

The only other input is cutoff_frequency (0.01–0.5, default 0.1), a normalized frequency. Lower = smoother output.

Output is filtered_sigmas, ready to feed onward just like any other SIGMAS.

Installing it

Bundled with the whole pack - ComfyUI Manager search "SigmaWaveFormNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/BenNarum/SigmaWaveFormNode

Restart ComfyUI. Dependencies are just numpy and torch, both already present; no models to download.

The footgun: leave the inverse Fourier toggle ON

apply_inverse_fourier defaults to true and you should keep it that way. Turn it off and the node hands you the raw frequency-domain coefficients instead of transforming back - a complex-valued spectrum, not a time-ordered noise schedule. PyTorch's conversion of that complex data to a float tensor just discards the imaginary part (on newer versions with a warning; on older ones it throws outright). Either way you get a meaningless sequence and no sampler is going to thank you for it. There's no realistic use case for the off position with this node - consider it a "don't touch" switch.

The other thing to remember is that filtering changes magnitudes. Zeroing out bins reshapes the schedule, and a lowpass with a very low cutoff can flatten your sigmas toward near-constant values, which defeats the point. Keep the cutoff high enough that the overall arc survives; you're removing jitter, not the curve.