ComfyUI Node
CNS Sampler Patch (Colored Noise)
CNS Sampler Patch: replaces white ancestral noise with wavelet-colored noise. Low-frequency noise is suppressed once those structures are built; high-frequency noise is boosted while fine details remain unresolved. Variance is strictly preserved per step (CNS energy budget constraint). Compatible with euler_ancestral, euler_ancestral_cfg_pp, dpm_2_ancestral, dpmpp_2s_ancestral, dpmpp_sde. Inert on deterministic (ODE) samplers.
CNS Sampler Patch (Colored Noise)
- sampler
- sampler
◄strength1.00►
◄gamma_power0.50►
◄gamma_scale2.0►
Categorysampling/custom_sampling/samplers
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| sampler | SAMPLER | — | |
| strength | FLOAT | 1.000–1 | [CNS] Effect strength. Lerp between white noise (0.0) and fully colored noise (1.0). Start at 1.0 and reduce only if artifacts appear. |
| gamma_power | FLOAT | 0.500.1–2 | [CNS] Exponent applied to the subband deficit before computing β. 0.5 (√) = CNS paper default — smooth, proportional routing. Higher (0.75–1.0) → sharper frequency routing, more aggressive. Lower (0.25–0.4) → softer routing, closer to white noise. |
| gamma_scale | FLOAT | 2.00.1–25 | [CNS] γ proxy divider. Scales down the subband energy ratios before computing deficits, preventing premature over-routing. Equivalent to 'γ Divider' in CNS paper (range 1.73–25.0). 1.0 = most aggressive (full energy ratio). 2.0–5.0 = recommended starting range for most models. Increase if early steps show structural instability. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| sampler | SAMPLER | — |