Latent Noise Injection

Latent Noise Injection takes a latent image and stirs some fresh noise into it. That sounds destructive, and used carelessly it is - but injecting a controlled dose of noise between sampling passes is a real technique for adding detail, breaking up over-smooth results, or giving a second sampler something to grab onto. It's a scalpel, not a hammer, and the whole game is dialing the amount.

The intuition: diffusion works by removing noise. If you hand a fresh KSampler a latent that's already fully "solved," it has little to do. Sprinkle a bit of noise back in and it re-engages - refining, hallucinating small detail, roughening up plastic-smooth areas. This is the manual, node-level cousin of what a second HR/refiner pass does automatically.

How it works

It generates noise and blends it into the latent you feed it, scaled by a strength value. Low strength nudges; high strength overwrites. The output is still a latent - same shape, same size - so it slots anywhere a latent goes, most usefully right before another sampling step.

A typical shape: KSampler → Latent Noise Injection → a second KSampler at low-ish denoise. The first pass builds the image, the injection re-roughens it, the second pass resolves that roughness into detail. You can also use it as a cheap variation trick - inject a touch of noise and rerun to shake a result loose from a local rut without changing the seed of your whole graph.

The inputs and outputs that matter

You feed it a latent (your samples) and set the noise strength - how much gets mixed in. That strength is the only real dial, and it's everything. The output is a latent you route into your next sampler or decode.

Keep the strength modest to start. A little goes a long way in latent space; what looks like "just a bit" can wreck the composition on the next pass.

How to install it

Comes with WAS Node Suite. Install the pack once - ComfyUI Manager, search WAS Node Suite, install, restart - or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui
pip install -r was-node-suite-comfyui/requirements.txt

then restart. On Windows portable, run the pip step with python_embeded\python.exe -s -m pip install -r ... (or the bundled install.bat). No models needed - this node is pure tensor math on your latent.

Common issues & troubleshooting

You injected noise and the second pass ignored it. If the following KSampler runs at very low denoise, it barely touches the latent, so your injected noise mostly rides through unchanged and shows up as grain instead of detail. Raise the second sampler's denoise a bit so it actually re-solves the noise you added - the two settings work as a pair.

The image fell apart. Too much strength. Injected noise scales fast; push it high and you don't get "more detail," you get a different, mushier image because you handed the next sampler enough noise to reinvent the composition. Back the strength down and creep up.

It's not a magic detail button. This adds entropy, not resolution. If you want an actually larger, sharper image, you want an upscale step (latent or pixel-space) - noise injection is for texture and variation, and it works best layered with an upscale/refine pass, not instead of one.

The suite won't import. WAS Node Suite is a large, unmaintained (since late 2023) pack whose signature failure is an "Import Failed" after a ComfyUI update, from a dependency version collision. Reinstall its requirements against your ComfyUI Python and restart. The node is fine; the suite install is the fragile part.