Get Sigma

Get Sigma doesn't touch a single pixel. It's a calculator: point it at your model plus your intended sampler, scheduler, and step count, and it tells you the exact sigma - the noise magnitude - a sampler would be working with at that point in the schedule. Its whole job is answering "how much noise do I actually need here," which is precisely the number Inject Noise otherwise makes you guess.

How it works

Sigmas are the real noise-level values a sampler steps down through as it denoises; the sampler decides how to remove noise at each step, the scheduler decides how much is there to remove, and Karras, exponential, and the rest of the scheduler list are all just different curves of these sigma values over your step range. Get Sigma runs your model/sampler/scheduler/step combination through that math and reads off the sigma at whichever point you specify - without generating or touching any actual image.

The inputs and outputs that matter

model, sampler_name, and scheduler need to match what your real sampling pipeline is actually using - this node has no way to know your setup except what you tell it. steps is your full schedule length. Most of the time start_at_step stays at 0 and end_at_step equals steps: that combination means "give me the noise level for a completely fresh generation." You only move them when you're splicing two samplers together - say one denoises steps 0 through 15 and a second continues from 10 through 20. Set start_at_step to 10 and end_at_step to 15, and you get the sigma for renoising the step-15 image back down to step 10 so the second sampler can pick it up cleanly. Output is a single FLOAT, and it plugs directly into Inject Noise's strength input - that's the entire point of this node existing.

How to install it

Manager search "ComfyUI Noise", or manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/BlenderNeko/ComfyUI_Noise

Restart afterward. No model downloads, no significant dependencies - it's math, not a runtime.

Common issues & troubleshooting

The most consequential mistake happens downstream, not on this node: forgetting to disable add_noise and return_with_leftover_noise on the Advanced KSampler that follows. If you're manually calculating and injecting noise, the sampler shouldn't be generating its own on top of it - both settings need to be off, or you're effectively dosing the latent twice, and results get muddy in a way that's genuinely confusing to debug if you don't know to check this first.

Second: sampler_name and scheduler here have to be an exact match for what your real sampler is running. Sigma curves differ meaningfully between, say, a Karras schedule and a straight one, so a mismatched pair hands you a number for a schedule you're not actually using - the injected strength will be wrong even though nothing looks broken.

And watch the defaults: this node's steps and end_at_step default to 10000, which is a ceiling the widget allows, not a sane starting value. Set both to your actual step count, or the sigma you get back has nothing to do with your workflow.