Dazzle KSampler Advanced (DazzleNodes)

Most samplers give you one eta slider and call it a day. Dazzle KSampler Advanced gives you two - plus separate controls for overshoot, SDE noise scaling, momentum, and the implicit-solver machinery underneath. It doesn't sample anything itself, though. This is the node you reach for when you want to rebuild the sampling math and hand the result to something else to run.

The one structural thing to understand up front: this node returns a SAMPLER object, not a LATENT. It's a factory. You wire its output into a node that actually consumes a SAMPLER - stock SamplerCustomAdvanced is the usual partner, and this pack's DazzleSharkSampler also takes one. That's the same pattern RES4LYF's Advanced node uses, which this is a port of.

Why you'd bother

The honest use case is fine-tuning on a flow-matching model where you've already accepted that sampler tuning is a real lever. The stock KSampler exposes one ancestral-noise knob; this exposes the whole RES4LYF tuning surface:

• eta / eta_substep - noise added then removed each step (and each sub-step, for samplers that use them). The sub-step pair only matters for RK methods with internal sub-steps.
• overshoot / overshoot_substep - boosts each denoising step's size then rescales; the tooltip describes it as a softening effect.
• s_noise - extra SDE noise; values around 1.03–1.07 give a moderate detail/paint-texture boost.
• d_noise - downscales the sigma schedule; 0.95–0.98 is a larger detail boost.
• noise_anchor - typically between 1.0 and 0.0; lower is grittier.
• momentum - accelerate convergence (positive when sampling, negative when unsampling).
• implicit_type / implicit_steps - selects the implicit solver family (rebound, retro-eta, bongmath, predictor-corrector) and how many implicit iterations to run. The fully implicit solvers (gauss-legendre, radau, lobatto) have no stock ComfyUI equivalent at all.

Defaults are sane - this is a "tune only what you understand" node. Every one of those dials starts at its conservative value, and the tooltips are unusually honest about which direction does what.

The inputs and outputs that matter

sampler_name (119 choices, default res_2m) and scheduler (default beta57) pick the solver and schedule exactly like the plain DazzleKSampler. latent_role is present but advisory - the tooltip says it plainly: this node doesn't consume a latent, and the actual noise-vs-init dispatch decision happens in whichever node eventually runs the SAMPLER. Don't expect it to do anything here.

The optional inputs are worth knowing: guides and automation are RES4LYF-ecosystem wiring, extra_options is a free-form string for advanced parameter injection, and options lets you merge in an existing OPTIONS object. The single output is sampler (SAMPLER), which you feed to SamplerCustomAdvanced or DazzleSharkSampler.

Installing and using it

Same pack, same install as the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/DazzleNodes/ComfyUI-DazzleKSampler.git

Restart ComfyUI and it's under RES4LYF/samplers. No extra pip packages, no model downloads. ComfyUI Manager users can just search "Dazzle KSampler".

Where people get burned

The number one mistake is expecting pixels out of this node - it returns a SAMPLER, not a latent, so wiring it straight to a VAE Decode just fails to connect. The second is turning every dial at once, which is how you end up debugging four interacting parameters on a random seed. Pick one (the community's most-tried is res_2m + beta57 + a nudge on s_noise or d_noise), lock your seed, and change one thing at a time. And a fair warning: this node is marked experimental in the code, and it's a thin, low-profile fork of RES4LYF's advanced sampler - great for tinkering, not the first thing you should hand to a beginner's workflow.