Advanced Noise Sampler (Nukun)

ComfyUI's advanced sampling path is powerful and spread out: you need a RandomNoise node, a SamplerCustomAdvanced, a scheduler, a guider - and if you want to control the noise's character, you're rebuilding the noise generator yourself. NukunAdvancedNoiseSampler bundles the noise generation and the sampling into one node, and it's the founding member of the Nukun sampler family: gaussian by default for a ComfyUI-compatible baseline, plus a shelf of alternate noise distributions for when plain Gaussian isn't giving you the variation you want.

The shape of the node

It takes the standard advanced-sampler inputs - guider, sampler, sigmas, latent_image - and handles the initial noise internally. The knobs that matter:

• noise_type - the star of the show. 20 distributions: gaussian (default), uniform, laplacian, pink, brown, blue, violet, pyramid, perlin, studentt, white, grey, velvet, plus the stronger experimental ones (green_test, rainbow_mild, rainbow_intense, wavelet, highres_pyramid, pyramid_discount5, pyramid_mix).
• noise_device - auto (default) generates noise on CPU like ComfyUI core does, for reproducible results; cuda uses CUDA when available and falls back to CPU. If you want your runs to match what a stock ComfyUI graph would produce, leave it on auto.
• noise_strength - a multiplier on the generated noise (1.0 = unchanged).
• add_noise - off means zero noise. This is your "sample from existing latent" mode.
• noise_seed - seeds the initial noise.
• preview_method - per-node preview control: default (follows ComfyUI's queue setting), latent2rgb (lightweight per-step preview), taesd (preview assets when available), none.

Outputs are the standard output, denoised_output, plus a seed output that exposes the final noise seed - handy for filenames and logging without hunting through the console.

The noise lore, distilled

The colored-noise spectrum (pink, brown, blue, violet) is about frequency character: pink/brown lean low-frequency (blobby, structural variation), blue/violet lean high-frequency (texture, grain), and pyramid/perlin give structured spatial patterns. These aren't magic settings - they change the character of the variation at the same strength. The Sonar-inspired profiles (green_test, rainbow_intense, velvet, wavelet) are implemented natively here - no ComfyUI-sonar dependency - and the README is blunt that they're stronger experimental profiles, easier to control in partial-denoise or multi-stage passes than in a full single pass.

How people actually use it

Three common setups:

1. Pure drop-in. gaussian + auto + strength 1.0 = ComfyUI-compatible noise. Wire it exactly where a RandomNoise+SamplerCustomAdvanced pair used to live and nothing changes.
2. Variation hunting. Your seed and prompt are locked, results feel samey, and you want more structural variety than Gaussian gives - try pink, pyramid, or highres_pyramid at strength 1.0 on an SDXL/Illustrious base.
3. Multi-stage passes. Advanced sampler workflows where the initial noise character matters more than it does in a single shot.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes.git

Restart ComfyUI or install via ComfyUI Manager ("Nukun"). The pack's runtime deps (numpy, Pillow, scipy, PyWavelets) cover the noise math; PyWavelets is what the wavelet profiles need.

The honest verdict

For most people, gaussian + auto + 1.0 is the whole node - and that's fine, because a compact ComfyUI-compatible advanced sampler with a seed output is already useful. The alternate noise types are a hobbyist toy until they're not: if you've hit the wall where every run on a fixed seed looks the same, colored-noise variation is a genuinely different dial from CFG or sampler choice. Start with pink or pyramid at low strength, and keep the wilder profiles for ranged passes where you can see what they're doing before committing a full run.