PreSampling (NoiseIn)

Straight from the pack author's own changelog: this node exists to replace an older one (easy latentNoisy) that had to sit awkwardly in front of your sampler to inject noise. preSamplingNoiseIn folds that same job into Easy-Use's normal presampling step instead, so noise injection becomes just another setting on the config node you're already using rather than an extra box cluttering your graph.

What noise injection is actually for

Injecting extra noise into a latent before sampling is a workaround for "this image is a bit too smooth/too settled" without doing a full denoise-from-scratch pass. It's the same idea behind SDE-style ancestral samplers adding stochasticity mid-run, except here you're doing it deliberately, once, up front, and dialing exactly how much. Common uses: breaking up an over-converged latent before a second pass, adding controlled variation to an otherwise-identical seed, or loosening up a latent that came from img2img and looks stiff.

How it fits the pipe pattern

Like every other Easy-Use pre node, this one doesn't sample - it configures. You feed it a pipe, set your sampling parameters and your noise factor, and it hands a configured pipe onward to whatever sampler node executes next in your chain. The noise gets mixed in as part of that later sampling step, not as a separate visible operation in the graph.

The inputs and outputs that matter

factor (0.1 default, 0–1 range) is the whole point of this node - how much extra noise gets mixed into the latent. Start low; 0.1–0.2 is a nudge, push toward 1.0 and you're basically doing a fresh generation with residual composition bias from the original latent. The rest of the required inputs are the same sampling settings you'd set on any Easy-Use presampler: steps, cfg, sampler_name, scheduler, denoise, and seed.

Two optional inputs matter if you're chaining this into something more complex: optional_noise_seed lets you decouple the noise-injection seed from your main sampling seed (useful when you want the injected variation to be independently reproducible), and optional_latent lets you feed in a specific latent to inject noise into rather than whatever the pipe is already carrying - handy when you're grabbing a latent from a batch index or a separate branch of the workflow.

Output is a single pipe, carried into your sampler node as usual.

Installing it

ComfyUI Manager: search ComfyUI Easy Use, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

Windows: double-click install.bat. Linux/Mac: cd in and run pip install -r requirements.txt - no script does it for you there. Restart ComfyUI.

Where people get stuck

The biggest mix-up is confusing this with a plain denoise adjustment. Turning factor up doesn't behave like lowering denoise on a fresh sampler - you're adding noise on top of an existing latent's structure, not controlling how much that structure survives resampling. If your results look identical to not using the node at all, check that factor isn't left at a near-zero default and that you're actually feeding the noised pipe into your sampler rather than a stale, unrelated one from earlier in the graph - an easy mistake once a workflow has more than one presampling branch. And if you're chasing reproducibility, remember optional_noise_seed exists specifically so a fixed main seed doesn't accidentally also lock your noise variation - leave it unset if you want the noise itself to vary between otherwise-identical runs.