Mix Noise with Latent

In any unsampling workflow - the logtd/ComfyUI-InversedNoise family, or any of the other "turn an image back into noise" techniques - one question dominates: how hard do you hold onto the original? Hold too tight and the re-sample just reconstructs the same image. Let go too much and you're back to ordinary img2img drift. Mix Noise with Latent is the dial for that decision.

What it does

MixNoiseNode takes a latent and a single float, randomness (default 0.1, range 0–1), and outputs a LATENT. That's the whole interface, and it's the only knob in the pack that has a slider. It blends your latent with fresh gaussian noise:

• randomness = 0 → the latent passes through unchanged.
• randomness = 1 → you get pure random noise.
• Everything between is a mix, with the default 0.1 meaning roughly 90% original signal, 10% fresh noise.

Here's the part people overlook: the mix is normalized - the result is divided by the combined magnitude of both components. So you're trading original noise for fresh noise without inflating the overall noise level. That's not a cosmetic detail. If it just added noise on top, every mix would make the latent louder and the inversion would smear; the normalization is what keeps the output usable as sampler input.

Where it actually goes

In the pack's example workflow it sits between the inverted-noise output of the SamplerCustom pass and the inject/sample step. Think about why: the noise recovered by inversion is approximate, and a small amount of fresh randomness breaks up the artifacts that approximate recovery leaves behind, while giving the re-sample a bit of creative slack. Crank it toward 0.5 and the structure starts to wash; push it high and you're doing plain img2img with extra steps.

You can also reach for it standalone, on any latent, as a poor-man's denoise controller - mix a little noise in, then sample. It won't match a real denoise-strength pass for control, but for adding variation to a latent it's dead simple.

Installing it

Same story as the rest of the pack: no requirements.txt, no model downloads, nothing beyond what ComfyUI already has. Via ComfyUI Manager search "InversedNoise", or:

cd ComfyUI/custom_nodes
git clone https://github.com/logtd/ComfyUI-InversedNoise
# restart ComfyUI

Gotchas

The trap is judging it by the wrong previews. The input is a latent - in an inversion graph you feed it the recovered noise, not the clean image latent, and the output only makes sense in that chain. Don't stare at the raw numbers or early previews; the 0.1 default is deliberately subtle, so expect changes to show up mostly in the final sampled image. And remember the whole pack is marked Experimental/WIP - treat the slider as "tune until it stops looking broken," not as a value with a canonical answer.