Universal Noise Sampler (Nukun)

Most samplers in ComfyUI draw their initial noise from one boring place: Gaussian. That's fine - it's what the models were trained on. But it's also why every attempt to vary starting texture meant bolting on a custom RandomNoise node and running everything through SamplerCustomAdvanced. Universal Noise Sampler (Nukun) collapses that whole rig into one node: it is the SamplerCustomAdvanced, it just lets you pick which noise recipe feeds it.

How it works

This is the explicit, graph-shaped sampler from the same personal pack as the Universal KSampler. Where that node builds everything internally, this one takes the pieces you assemble yourself: a GUIDER (from a CFGGuider node), a SAMPLER (from KSamplerSelect), SIGMAS (from a BasicScheduler or similar), and your latent_image. The README's recommended wiring is exactly that: CFGGuider + KSamplerSelect + BasicScheduler in, latent out.

Inside, it's a drop-in replacement for the RandomNoise + SamplerCustomAdvanced pair. The pack's own noise generator handles noise_seed, noise_device, and noise_profile, then feeds the sampled latent through the guider and sampler you supplied.

The inputs that matter

• guider, sampler, sigmas, latent_image - your standard advanced-sampler plumbing.
• noise_profile - the star. 29 choices spanning basic spectral types (gaussian, uniform, laplacian, pink, brown, blue, violet, pyramid, perlin, studentt, white, grey, velvet, wavelet, the experimental green_test / rainbow_* / highres_pyramid set) plus composite recipes tuned for Illustrious and Pony v7.
• noise_seed - used for the profile, with the usual control-after-generate increment/randomize behavior.
• noise_strength - scales whatever profile you picked. Start low on the wild ones.
• detail_bias - composite profiles only; lower pushes toward big forms, higher toward micro-detail.
• noise_device - auto keeps CPU noise for reproducibility, matching ComfyUI-core.
• add_noise - disable for a zero-noise pass (the start of a multi-stage handoff, or a purely denoised img2img run).

Outputs are output, denoised_output, and seed - the seed output is genuinely useful if you want the exact noise seed in a filename without reconstructing it from the widget.

The honest take

The gaussian baseline (gaussian + auto + 1.0) is a faithful ComfyUI-core reproduction, so this node costs you nothing to try. The extra profiles are where it gets opinionated - colored noise changes composition and texture, and people react to it the way they react to ancestral samplers: sometimes magical, sometimes a mess. Treat pink/pyramid as the tasteful end and green_test/rainbow_intense as "I'm bored" territory. When a workflow goes off the rails after a profile swap, the noise is doing it - reset to gaussian before you blame your CFG or scheduler.

Installing it

It ships in the Nukun_ComfyUI_Nodes pack. Install via ComfyUI Manager (search "Nukun") or:

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

Restart ComfyUI. Dependencies are just numpy, Pillow, scipy, and PyWavelets - nothing heavy, no models to download. If you want a KSampler-shaped version of this same node instead, use Universal KSampler (Nukun) from the same pack.