Nodes/SigmaWaveFormNodes/Attenuator Node
ComfyUI Node

Attenuator Node

The volume knob for sigma schedules

By BenNarum·Created 2 years ago·Updated 2 years ago· 6
Attenuator Node
  • sigmas
  • attenuated_sigmas
attenuation_factor1.0

The Attenuator Node is the simplest node in the SigmaWaveFormNodes pack, and it earns its place by being the one you'll actually reach for. It does exactly one thing: multiply a SIGMAS sequence by a single factor. In a pack of experimental wave-drawing toys, this is the boring utility piece that makes the toys usable.

Here's why it matters. The waveform generators in this pack produce sigmas with whatever magnitudes the math happened to output - a square wave at amplitude 3, a random wave scaled to its own range - and none of it is guaranteed to sit anywhere near what a sampler expects. You generate a wave, it comes out too hot or too weak, and you need a trim before it hits the sampler. That's this node. The README itself frames it as the fine-tuning stage: use it after a waveform generator, or after a Fourier filter has reshaped the sequence.

What it does

One input sequence, one float, one output. Specifically:

  • sigmas - any SIGMAS sequence, from this pack or anywhere else.
  • attenuation_factor - default 1.0, range −10 to 10. Multiply the whole sequence by this.
  • Output: attenuated_sigmas.

A factor below 1 shrinks the whole schedule - gentler denoising, less aggressive noise removal per step. Above 1 amplifies it. There's no clamping, so big factors push sigma magnitudes well outside normal territory; a factor of 10 is not a subtle adjustment, it's a demolition order.

Installing it

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

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

Restart ComfyUI. Deps are numpy and torch only; no models, no downloads.

Notes and traps

The range goes negative, and a negative factor doesn't just quiet the schedule - it flips it upside down, turning high noise into low and low into high. There's a tiny chance that's what an experiment calls for, but treat negative values as a deliberate inversion, not a louder version of zero.

One thing the node does not do is restore monotonicity. If the input wave oscillates, the attenuated output oscillates - scaled, but still not a proper descending schedule. The Attenuator is a trim, not a fix. Pair it with damping and the Fourier filter for that. And since it's a plain element-wise multiply with zero state, you can chain a few of them without worry - though if you need more than one, you're probably fighting the generator instead of tuning it.

Categorysampling/custom_sampling/attenuators

Inputs (2)

NameTypeDefaultDescription
sigmasSIGMAS
attenuation_factorFLOAT1.0-10–10

Outputs (1)

NameTypeDescription
attenuated_sigmasSIGMAS