Flex Latent Noise ⚡🅡🅞🅣🅘

Usually you're trying to get noise out of a latent - that's what the sampler is for. Flex Latent Noise does the opposite on purpose: it injects a controlled amount of noise back in. That's a legitimate technique, not a mistake - a touch of reintroduced noise before a low-denoise pass can add texture and grain a completely clean latent won't have on its own, and because this is a Flex node, the amount of noise can react to an external signal instead of sitting at one fixed value the whole clip.

How it works

noise_type picks the distribution: Gaussian is the standard bell-curve noise diffusion models are actually trained around (the natural choice most of the time), and Uniform spreads values evenly across the range instead, which reads as flatter and grainier rather than "soft." noise_level sets how much gets mixed in. Feed it a FEATURE and set feature_param to noise_level, and the amount of noise can pulse with whatever's driving it - audio energy is the obvious one, giving you a latent that gets grainier or shimmers more intensely on a beat, with feature_mode controlling whether that's a push around your base level (relative) or a scale from zero up to it (absolute).

Inputs and outputs that matter

• latents (required, LATENT) - the latent to add noise to.
• noise_type - Gaussian or Uniform.
• noise_level (default 0.1, 0–1) - how much noise gets mixed in.
• opt_feature (optional, FEATURE) - wire this in to make noise_level react over time.
• Output - a single LATENT.

Installing it

Via ComfyUI Manager: search "RyanOnTheInside," install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/ryanontheinside/ComfyUI_RyanOnTheInside.git
cd ComfyUI_RyanOnTheInside
pip install -r requirements.txt

then restart ComfyUI. No models to download - it's tensor math, ready as soon as the dependencies are in.

Common issues & troubleshooting

Output looks broken or noisy after the sampler. noise_level is deceptively powerful - even modest values reintroduced right before a low-denoise pass can dominate the result, since the sampler has less room to "clean it up." Start well under the default and increase gradually rather than dialing it in from a high value downward.

No visible difference at all. Check where in your graph this sits relative to your sampler. Noise injected into a latent that then goes through a high-denoise pass gets mostly overwritten by the sampling process itself - this node reads best on a latent headed into a light, detail-preserving pass, not a from-scratch generation.

Feature-driven noise looks erratic rather than musical. A raw, unsmoothed audio feature will produce jumpy, inconsistent grain frame to frame. If you want the shimmer to read as pulsing rather than flickering, smooth the feature upstream before it reaches opt_feature here.