Sampler Settings

Sampler Settings is the pack's version of the KSampler's noise widgets, packaged as a bundle so a single node can carry a complete recipe: seed, steps, cfg, sampler, scheduler, denoise, and a few advanced knobs the stock KSampler doesn't even show. By itself it does nothing but hold settings - you wire its sampler_settings output into Ouroboros, Phantas, Chimera, or Siren (Music Sampler). What makes it worth learning is the chaining: wire one node's output into the next one's sampler_settings input and you've built a multi-stage refine pipeline that all those samplers understand.

The dials that matter

seed plus seed_mode is the interesting part. The default fixed keeps the same seed every iteration, which is what makes Ouroboros's self-correcting loop mean anything - only the prompt varies, so score changes are about the writing, not the noise draw. random explores but "mixes 'the prompt improved' with 'the seed got lucky'," as the tooltip bluntly puts it; increment/decrement step the seed by seed_step each iteration. For a one-shot render (no loop), this is all inert - it matters only inside an iterative sampler.

The advanced knobs are the "reach the full sampler API" layer, and they only do anything on samplers that support them:

• eta - stochasticity for ancestral/SDE samplers (euler_ancestral, dpmpp_*_sde). 0 = deterministic ODE. No effect on deterministic samplers like euler.
• s_noise - multiplier on the noise added at stochastic steps. >1 adds grain/detail, <1 smoother.
• s_churn - re-injects noise into otherwise-deterministic samplers (euler, heun, dpm_2). 0 = off. Small values add detail; too high adds artifacts.
• solver_type - has two separate vocabularies: midpoint/heun for the dpmpp_2m_sde family, phi_1/phi_2 for the SEEDS and exp_heun families. A value from the wrong family is ignored rather than breaking the run, so you can't hurt yourself.

Every knob is only passed to samplers that actually accept it, and defaults keep a single stage bit-identical to the classic KSampler path. So you can treat this node as "a KSampler I can save and chain."

Chaining into refine mode

Feed two or more chained Sampler Settings nodes into Ouroboros (or Phantas, which takes the same bundle) and it runs in refine mode: stage 1 drafts at full denoise, later stages polish the previous latent - give them denoise < 1 - and in Ouroboros the critic judges only the final image. Left→right is stage order. One node = classic single pass, fully backward compatible. denoise is also how you do img2img: feed a real latent into Ouroboros and set denoise < 1 here.

Installing and using

It's part of Kinburg/Kinburg-Nodes - ComfyUI Manager search "Kinburg-Nodes", or cd ComfyUI/custom_nodes && git clone https://github.com/Kinburg/Kinburg-Nodes, restart. No dependencies, no models. Don't hunt for image size here - it comes from the latent wired into the sampler, not the bundle.

One habit worth forming: wire a Settings Save between a chain and its sampler with a preset_name filled in, and the recipe (with score and seconds) lands in the model's preset library, where Settings Select can pull it back out. That's the whole point of a bundle - it turns "what settings did I use for that good one?" from a memory problem into a lookup.