Universal KSampler (Nukun)

You've got a working KSampler and you're happy with it. The last thing you want is to rewire the whole sampling section just to try a different kind of noise. Universal KSampler (Nukun) is the node that lets you skip the rewiring: it looks and behaves like a stock KSampler, but the noise inside is a Nukun noise profile instead of plain Gaussian.

It's one node in a small personal pack by OnekoSL (Nukun_ComfyUI_Nodes). Don't come here expecting a huge community footprint - it's a utility drawer, and this is the drawer's friendliest knob.

How it works

A normal KSampler hides the machinery: it builds a CFG guider from your model and conditionings, picks the sampler object, computes a sigma schedule, and generates initial noise. Nukun Universal KSampler does all of that internally too - reading the source, it literally constructs CFGGuider, the sampler, and the sigmas for you. The difference is the noise step. Instead of calling ComfyUI's stock random noise, it hands your noise_profile to Nukun's own generator.

That's the whole pitch: you keep your KSampler-shaped graph, and gaussian + auto + 1.0 is a drop-in, ComfyUI-core-like baseline. Change nothing, get the same result as a stock KSampler. Then, when you're curious, flip noise_profile to pink, pyramid, or one of the fancier composites and rerun on the same seed to see how the starting texture changes the image. These expanded profiles are Sonar-inspired but implemented right in the pack - no separate ComfyUI-sonar install needed.

The inputs that actually matter

Most of the inputs are the ones you already know from KSampler: model, positive, negative, latent_image, steps, cfg, sampler_name, scheduler, denoise, seed. The interesting ones are:

• noise_profile - 29 options. Basic types like gaussian, uniform, pink, brown, blue, violet, pyramid, perlin, plus composite illustrious_* and pony_v7_* recipes.
• noise_strength - a multiplier on the generated noise. 1.0 is the faithful default; strong experimental profiles like green_test or rainbow_intense are much easier to control at lower strengths.
• detail_bias - only affects composite profiles. Lower emphasizes larger forms, higher emphasizes detail.
• noise_device - auto uses CPU for reproducibility, cuda if you want GPU noise (falls back to CPU).
• add_noise - uncheck to sample with zero noise.

Three outputs come out: output (the latent you VAE-decode), denoised_output (the direct denoise result, handy for comparisons), and seed - the actual noise seed used, which is nice for filenames.

One honest caution

cfg defaults to 7, which is an SD 1.5/SDXL number. If you're sampling a flow-matching model (Flux, Wan, Z-Image), that default will burn your image - check what your model actually wants before you blame the noise profile.

Installing it

Easiest is ComfyUI Manager - search "Nukun". Or clone it:

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

Restart ComfyUI. Runtime deps are just numpy, Pillow, scipy, and PyWavelets - no model downloads for this node.

If your output suddenly looks different after switching from a stock KSampler, your noise_profile or noise_strength is doing something. gaussian + auto + 1.0 is always the "did I break it?" reset.