Plasma Noise

This is the node the pack is named after, and it's the one worth building a workflow around. JDC_Plasma generates "extremely noisy fractal diamond square noise clouds" - which is a fancy way of saying it produces those swirling, colorful, infinitely-smooth-but-random images you remember from every 90s demo scene. It's random noise, but structured random noise. And that structure is exactly why it's more interesting than a sheet of static.

It's one of the five noise generators in comfy-plasma by Jordach, and the reason the pack exists at all. The community has actually used these nodes in real workflows - they showed up in the notable SD3 layer-noise-injection experiments people were running when the model was new.

How it works

Under the hood it's the classic diamond-square midpoint displacement algorithm, run three times - once per RGB channel. You seed four corner values, then recursively subdivide, averaging neighbors and adding a random displacement that scales with distance. The result is a fractal: locally chaotic, globally coherent, like clouds or terrain.

That fractal character is controlled by turbulence (a float from 0.5 to 32, default 2.75):

• Lower turbulence (near 0.5) - smooth, large, blobby clouds. The displacement is small, so neighbors stay close.
• Higher turbulence (up to 32) - more displacement per step, so the result disintegrates into something closer to static.

The README describes it perfectly: low values give "smoother, larger clouds," high values give "more static-like noise." Tune it to taste and you'll see why people keep coming back to this node - there's a whole spectrum between "smooth marble" and "broken TV" hiding in that one slider.

Like the other generators, it also exposes value_min/value_max and per-channel red/green/blue_min/max range controls. Default -1 means full range (0–255), and per-channel values override the global ones. So you can bias a plasma toward cyan, clamp the brightness, whatever.

Why you'd actually use it

The practical trick - and the one the author's own example workflow demonstrates - is to VAE-encode the plasma into a latent and feed it to the sampler as your starting point instead of plain latent noise. Structured noise in, structured-ish image out. Combined with the pack's JDC_PlasmaSampler (which has a latent_noise dial for exactly this), it's a way to nudge img2img-style generation toward the shapes you've drawn with the plasma. People have also used it as a texture source and as a masking pattern. It's a creative toy with a real job.

Inputs & install

The inputs that matter: width/height (128–8192, steps of 8), turbulence, the range clamps, and seed. Output is a single IMAGE.

Install is painless - no requirements.txt, no models:

ComfyUI Manager → Install Custom Nodes → search comfy-plasma → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Jordach/comfy-plasma

Restart ComfyUI. You'll find it under image/noise.

Gotchas

Same seed caveat as the other noise nodes: the seed is reproducible, but the author notes ComfyUI treats these as static-output nodes, so you have to actually change an input to force a fresh roll. And diamond-square recursion at very high resolutions is genuinely slow - each channel does a full subdivide pass in Python. Generate at 512, upscale if you need bigger. There's also a quirk where the node makes its working canvas square (the larger of width/height) and crops to your requested size, so odd aspect ratios can feel slightly off - keep it square-ish and you won't notice.