SamplerSonarEulerA

If you ask the author of ComfyUI-sonar which momentum sampler to start with, this is the one. "I personally recommend using the custom sampler approach and the ancestral version," the README says, and this node is that recommendation in node form: ancestral Euler with a momentum history layered on top.

Ancestral samplers - Euler a, DPM++ SDE - add a little fresh noise at each step, which is why they never quite settle the way converging samplers do. Same seed, more steps, and the image keeps drifting instead of refining. That's a feature when you want variety and character, which is most of the time when you're generating for fun rather than debugging. Momentum sampling adds a running history of where the denoising direction has been pointing, then blends it in each step, which smooths out some of the chaos that pure ancestral Euler gets up to.

The inputs that matter

It has the same core momentum controls as its plain sibling (momentum, momentum_hist, momentum_init, direction, rand_init_noise_type - same defaults of 0.95 / 0.75, and momentum = 1.0 disables momentum entirely), plus the two ancestral knobs:

• s_noise (default 1) - multiplier for the noise added during ancestral sampling. Turn it down if output gets grainy at higher steps.
• eta (default 1) - basically controls how ancestral the sampler is. 0 gives you a non-ancestral (or SDE) sampler.
• noise_type (default gaussian) - which noise family feeds those per-step ancestral injections. This is where the pack's noise toybox starts to matter: switch to brownian or a pyramid variant and the character of the image changes even though the sampler math doesn't.

There's also an optional guidance_cfg_opt socket for extended guidance parameters, and an optional custom_noise_opt if you build a custom noise chain - when that's connected it overrides noise_type entirely.

How to wire it

The node outputs a SAMPLER socket, so it goes into the custom sampling pipeline: SamplerCustomAdvanced + BasicScheduler, or a CFGGuider into SamplerCustom. If you want momentum active only part of the way through a generation, or you want the denoised momentum mode, hang a SamplerConfigOverride off the chain - you can pass YAML to set momentum_start_step and momentum_end_step, or switch momentum_mode between classic, new, and denoised.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/blepping/ComfyUI-sonar

or search "ComfyUI-sonar" in ComfyUI Manager and restart. No models to download. The same pack-level caveats apply everywhere in this repo: inputs change between revisions (track your git hash if you want reproducible gens), and seeds aren't guaranteed portable across GPUs because some noise generation happens on CPU.

Where people get burned

The predictable ancestral trap is expecting determinism. If you're A/B testing a setting change, use a converging sampler instead - this one will keep throwing you off with noise variance. And if output ever looks mushy, s_noise is the first dial to try, before you go hunting in the momentum knobs.