CustomNoise

ComfyUI's SamplerCustomAdvanced node is picky about what it accepts for noise - it wants a NOISE-typed input, normally supplied by RandomNoise or DisableNoise, not a raw latent. CustomNoise is a small but useful adapter: it takes any ordinary LATENT and wraps it as NOISE, so you can use literally anything as your starting noise instead of whatever the built-in noise nodes generate for you.

Why you'd want this

The default noise sources give you either fresh random noise or none at all - reasonable defaults, but limiting if you want actual control over what the sampler starts from. With CustomNoise, "noise" can be anything you can produce as a latent: noise you built with specific per-channel statistics (this pack's own Offset Latent Image and RandnLikeLatent nodes are built for exactly that), a latent saved from a previous run for exact reproducibility, or something more experimental - deliberately structured "noise" for creative effect rather than pure randomness. Anywhere a workflow calls for controlling the sampler's starting point precisely rather than trusting a random seed, this is the node that makes it possible.

It's a genuinely tiny node - a type adapter, nothing more - and that's the whole appeal. SamplerCustomAdvanced's ecosystem of noise nodes is built around the assumption that you either want randomness or you don't, and CustomNoise is the escape hatch for everyone in between: anyone who has a specific latent in mind and just needs ComfyUI to stop complaining about the type mismatch.

Inputs and outputs that matter

There's exactly one input and one output:

• noise - the LATENT you want to use as your starting noise.

Output: NOISE, wired directly into SamplerCustomAdvanced's noise input in place of RandomNoise/DisableNoise.

Installing it

Through ComfyUI Manager, search ComfyUI Image Filters, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/spacepxl/ComfyUI-Image-Filters

Restart. No models to fetch - it's a thin type wrapper, effectively free to run. The pack-wide opencv dependency can still bite on install if another node pack has a conflicting version; run the repo's import_error_install.bat, or manually reinstall just opencv-contrib-python after removing the other opencv variants.

Where people get burned

This node only makes sense alongside SamplerCustomAdvanced - it produces a NOISE-typed output, which a plain KSampler won't accept as an input at all. If you're not already using the custom/advanced sampling nodes, you don't need this one.

The other thing worth knowing: whatever latent you feed in becomes the entire starting noise, not an addition on top of random noise. If your latent isn't actually noise-shaped - say, you accidentally wired in a real image's encoded latent instead of an actual noise tensor - the sampler will still run, but the result won't look like a normal generation, because the sampling process assumes it's starting from something statistically noise-like. If your results look strange in a way that doesn't map to any of your normal settings, check that what's actually flowing into this node is noise, not a rendered latent.