SamplersSUNoiseAdvanced

SamplersSUNoise is the version you use to try the pack. This is the version you use once you've decided you like it and want to squeeze it. Same seven ancestral samplers, same SUNoise mechanism, but now eta, s_noise, solver_type and r are exposed instead of being silently fixed at the author's defaults. If you've ever wished ComfyUI's stock samplers let you reach the k-diffusion parameters underneath, this is that.

The knobs, in the order you'll actually touch them

• s_noise (default 1.0) - scales how much fresh noise gets injected at each step. This is the one most people tune. Crank it up (2–4) for more variety and "looser" results from your ancestral sampler; drop it toward 0 and the sampler stops being ancestral and starts behaving like a converging one - same seed, same image, every time. It's the classic s_noise from k-diffusion, just wired to SUNoise.
• eta (default 1.0) - controls the size of the ancestral jump itself: how far ahead the sampler looks before re-injecting noise. Higher eta = more stochastic variation; eta at 0 turns the stochastic part off entirely (at which point you're basically running a plain ODE solver and wondering why you installed this pack).
• solver_type - midpoint or heun. This only does anything when you've picked dpmpp_2m_sde; it picks the correction term for the SDE's second-order step. Heun is the default and generally smoother; midpoint is slightly cheaper. If you can hear the difference on your images, you have better ears than most of us.
• r (default 0.5) - only used by dpmpp_sde. In that solver the noise is normally drawn from a Brownian tree sampler, and r sets where in the step the noise is sampled. It's a fine-tuning knob for the stochastic SDE path; 0.5 is the sensible starting point.

The parameters are per-node, not per-sampler - all four are always present, but as the list above says, only the relevant ones do anything for the sampler you've selected. The extra knobs are ignored for the others, so you don't need to "reset" them when you switch samplers.

What wires where

Output is a single SAMPLER that plugs into SamplerCustomAdvanced's sampler input. Same wiring as the non-advanced node and every other sampler node in this pack - model → sampler → sigmas → latent. The noise_type choice (standard vs multires) behaves exactly as in the plain node: uniform noise per pixel, or uniform noise built at two resolutions and summed for low-frequency structure.

Install is identical to the rest of the pack - ComfyUI Manager, search "SUNoise", or:

cd ComfyUI/custom_nodes
git clone https://github.com/bvhari/ComfyUI_SUNoise

Restart, done. No dependencies beyond what ComfyUI ships, no model downloads, GPL-3.0.

The honest take

Is the Advanced node worth it over the plain one? If you're not going to touch the knobs, no - the plain SamplersSUNoise uses the same defaults (eta 1.0, s_noise 1.0, solver_type heun, r 0.5) that this one starts from. The Advanced version earns its name only when you want to fight the samplers' behavior: s_noise is genuinely useful for dialing how "ancestral" your output feels, and eta is how you push variation further than the defaults allow.

Where people get burned: forgetting that ancestral samplers never converge, then raising s_noise and eta together and getting images that wander further and further from the prompt. If that's you, keep eta modest and raise s_noise alone. And remember the README's standing advice: these samplers want 30 steps or more. At 10 steps you're changing the noise distribution at every step of a run that was already too short to settle.

One more thing to know: this is a small pack with essentially no community chatter around it, so most of what you'll learn about the knobs will be from your own sweeps. Fix the seed, vary s_noise, and see what your checkpoint actually likes.