Colored Noise Sampler

Colored Noise Sampler is the heart of akashzeno/ComfyUI-ColoredNoiseDiffusionSampling. It outputs a SAMPLER - a drop-in for the sampler slot in SamplerCustom or SamplerCustomAdvanced - that wraps your usual stochastic sampler and colors the noise it injects at every step. The all-in-one ColoredNoise_KSampler is this same machinery in a KSampler-shaped wrapper; this node is the raw version, for when you're already building the custom-sampling graph.

The idea comes from the Colored Noise Diffusion Sampling paper (Davidson, Issachar & Benaim, arXiv:2605.30332): standard SDE samplers pump white noise into every step, wasting energy on frequency bands that are already resolved. Colored noise routes that energy toward the bands that still need work - low-frequency structure early, high-frequency detail late. Here's how this pack does it: draw white noise, FFT it, multiply the spectrum by a radially symmetric per-frequency profile, inverse-FFT, renormalize to unit variance. The sampler's own sigma_up · s_noise amplitude still applies on top; the pack just changes the flavor of what gets injected.

What you actually set

• base_sampler - dropdown is pre-filtered to stochastic samplers only (ancestral/SDE like dpmpp_2m_sde, plus the _RF rectified-flow variants for Flux/SD3). Deterministic samplers never inject per-step noise, so coloring them would do nothing - the pack doesn't even list them.
• mode - parametric (default) or gamma_matrix. Start parametric: it's a model-agnostic power-law profile amplitude(f) ∝ f^(−α/2) with the exponent interpolated from alpha_start (high sigma) to alpha_end (low sigma), linearly or exponentially.
• eta and s_noise - the universal SDE knobs, forwarded to the base sampler. eta = 0 makes the step deterministic and disables colored noise entirely, so leave it alone unless you're deliberately testing.
• energy_scale - post-renorm "heat" knob. Default 1.0 is neutral; anything else intentionally changes the noise level.

The gamma controls (gamma_matrix, gamma_divider, gamma_shaping, power_gamma, alpha_tilting) only matter in gamma_matrix mode - the paper-faithful path that loads a [steps, bins] matrix from models/colored_noise_gamma/. Two SiT/ImageNet matrices ship with the pack. Honest warning, straight from the README: on SD/SDXL/Flux those matrices are a documented heuristic, not a calibrated schedule, so use parametric for real work and gamma_matrix for experimenting.

How to wire it

Colored Noise (Initial) ─┐
Colored Noise Sampler ───┤
BasicScheduler ──────────┼─► SamplerCustomAdvanced ─► VAE Decode
BasicGuider ─────────────┘

This node colors per-step noise; the initial latent start is a separate concern, which is exactly why the pack ships ColoredNoise_Noise for the noise input.

Install and sanity checks

No pip dependencies (torch ships with ComfyUI). ComfyUI Manager → search "Colored Noise", or:

cd ComfyUI/custom_nodes
git clone https://github.com/akashzeno/ComfyUI-ColoredNoiseDiffusionSampling

Restart, and confirm the startup log: [ColoredNoiseDiffusionSampling] loaded: 3 nodes | 22 stochastic base samplers | .... Each generation logs which base sampler and alpha sweep it ran, so you can verify coloring is active rather than guessing.

One note on expectations: the paper's quantitative wins were measured on SiT, not on the models most of us run. This is a quality/character knob - some models genuinely like the broadband-to-detail sweep, others shrug. Cheap to test, worth having in the box.