ClownSamplerAdvanced

This is ClownSampler with the hood off and every wire exposed. Like the regular one, it doesn't sample anything itself - its only output is a SAMPLER object that you plug into a node that runs it, like SharkSampler or ComfyUI's SamplerCustomAdvanced. What the "Advanced" buys you is control: separate noise settings for steps and substeps, overshoot, momentum, implicit solver refinement, and the guides input, all laid out as individual knobs instead of hidden inside a convenience node.

You reach for this when the everyday sampler isn't enough and you know why you're here - dialing in SDE noise behavior on a substep sampler, adding implicit refinement steps for coherence on SD3.5 Medium, or attaching composition/style guides directly to the solver. It's a power tool. If you don't yet have a specific reason to touch overshoot or implicit steps, the plain ClownSampler or the all-in-one ClownsharKSampler will serve you better.

How it works

It assembles a sampler definition from a long list of parameters and outputs it as a SAMPLER. The key idea running through the knobs is the step/substep split: solvers ending in s (res_2s, res_3s) take substeps between steps, and this node lets you set noise type, eta, and overshoot separately for the main steps and those substeps. Implicit steps add iterative refinement on top. Nothing renders here - it's the recipe, cooked elsewhere.

The inputs and outputs that matter

Out of the long list, the ones a beginner should actually touch:

• sampler_name (default multistep/res_2m) - the solver. res_2m is the everyday pick: one model call per step, works on almost everything. The s solvers are slower but more accurate.
• bongmath (BOOLEAN, default true) - leave it on. It aligns each substep's latents with the noise predictions as it goes, effectively denoising forwards and backwards at once, for a more accurate result with no extra VRAM and no extra time. Set-and-forget quality.
• eta (FLOAT, default 0.5) - how much noise the SDE-mode solver re-injects per step. eta_substep does the same for substeps.
• implicit_steps (INT, default 0) - iterative refinement passes; each one multiplies runtime, and gains fade after 2–3. Leave at 0 unless you're fighting artifacts on something like SD3.5 Medium.

Optional inputs include guides (RES4LYF composition/style guides), automation, and an extra_options text box for advanced flags. The single output is sampler (SAMPLER) - run it to a SharkSampler or SamplerCustomAdvanced.

Installing it

ComfyUI Manager: search RES4LYF, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF/
cd RES4LYF
pip install -r requirements.txt

pip install in the venv or portable Python, restart, hard-refresh (F5). Install rgthree-comfy too - the nested multistep/res_2m menu wants its "Auto Nest Subdirectories" setting on.

Where people get tripped up

The first shock is that nothing happens. This node produces a SAMPLER, not an image - it must feed a node that supplies the model, latent, and sigmas. On its own it renders nothing, and that's by design.

The second is analysis paralysis. There are dozens of floats here, most defaulting to 0 or 1 for a reason: they're identity values, and you can safely ignore the ones you don't understand. Don't crank overshoot, momentum, and noise_boost all at once and expect sense - change one thing, keep your seed fixed, compare. And a real cost worth naming: implicit_steps and the s-family solvers multiply your model calls, so a "small" change can triple your render time. Know which knobs cost compute before you turn them.