😼> Noise From Image

This one's a bit different from the rest of the pack. Where most of ComfyUI_yanc is small, utilitarian glue nodes, Noise From Image is the technical half of a specific technique the pack's author (who runs the "A Latent Place" ComfyUI YouTube channel) has been teaching since at least 2023: injecting hand-crafted noise into the sampling process, on top of the usual random init noise, to push SDXL and similar models toward more detail and texture than a clean seed alone gives you. It caught enough attention early on that other people were remixing his noise-injection workflow within days of the original tutorial going up.

The node generates a noise pattern derived from an input image's structure and blend settings, and outputs it both as a viewable image and as a LATENT - the latent is what actually gets fed into the noise-injection sampling chain (paired with YANC's NIKSampler) rather than a plain KSampler's default noise.

How it works

Under the hood it's building layered noise: a base "waviness" pattern (magnitude/smoothness), plus a separate finer noise-dot layer (noise_intensity, noise_resize_factor) blended on top at a controllable rate. If you feed it a batch of images, blend_mode and blend_rate let it progressively blend each generated noise pass with the one before it rather than treating every batch image as independent - off by default, since most people are just injecting noise for a single image.

The inputs and outputs that matter

The parameters are genuinely a lot, but only a couple actually need touching for a first pass - the rest are fine-tuning once you know what you're chasing:

• magnitude (default 210, 0–250) - the author's own description: "the waviness of the noise." This is the main knob for how aggressive the injected noise looks.
• smoothness (default 3, 0–10) - lower values give a grainier result; higher, a smoother one.
• noise_blend_rate (default 0, 0–1) - controls how much of the extra noise-dot layer actually gets applied; the author's guidance is 0.15–0.2 is a good starting value, and at 0 the extra noise layer is skipped entirely, which is worth knowing if you set everything else and see no effect.
• saturation_correction (default 1, 0–1.5) - if your output colors look washed out after injecting noise, raise this.
• vae_opt (optional VAE) - needed to actually produce the latent output; without a VAE connected you'll only get the visual image preview, not a usable latent to sample from.

Outputs: image (a preview of the generated noise) and latent (LATENT) - the one you actually route into the noise-injection sampling chain alongside YANC's NIKSampler.

How to install it

Via ComfyUI Manager: search "yanc" or "YANC - Yet Another Node Collection," install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ALatentPlace/ComfyUI_yanc

then restart. No extra Python dependencies to install and nothing to download - it's tensor math on top of what ComfyUI already ships.

Common issues & troubleshooting

The latent output is empty or the node errors. Check vae_opt is actually connected - it's optional in the schema, but it's required to produce a real latent output rather than just an image preview.

You're getting no visible effect from the extra noise layer no matter what you set. noise_blend_rate defaults to 0, which means that layer contributes nothing until you raise it - this is the most common "why isn't this doing anything" trap with the node, since every other parameter still runs even at rate 0.

Colors look off after injection. That's what saturation_correction is for - the README specifically calls out washed-out color as the symptom to fix by raising it, so don't reach for a separate color-correction node first.

You're not sure this is worth the complexity. Fair - it's a niche, stylistic technique aimed at squeezing extra perceived detail/texture out of SDXL-era models, not a correctness fix. If you're not already chasing that specific look, the defaults with a plain KSampler will serve you fine without the extra graph.