Legacy2_ClownsharKSamplerAutomation

A stated goal of RES4LYF, in the author's own words, is "to facilitate modulating parameters vs. time." This is the node that does it. Normally your sampler settings are one number each - eta is 0.5 for the whole run, s_noise is 1.0 for the whole run. Automation lets you make them curves instead: a lot of noise early and none late, or the reverse, or whatever shape you want across the step count.

That sounds fussy, and it kind of is, but it's also where a chunk of RES4LYF's quality gains come from. Adding controlled noise in the early steps and pulling it back as the image resolves is a real technique, not a gimmick - the trajectory-shaping the pack is built around. This node is how you express that shape without hand-editing anything inside the sampler.

How it works

It packages up per-step schedules and hands them to the sampler as one AUTOMATION object. Inside ClownSamplerAdvanced there's an optional automation input - you plug this node's output straight into it, and from then on the sampler reads its eta and noise values from your curves instead of the single widget values.

The curves themselves come in as SIGMAS - the same list-of-numbers type that scheduler and sigma nodes emit. So you build a curve with whatever sigma-generating node you like (a plot node, a schedule node, a float-list) and route it here.

The inputs and output

Everything on this node is optional, and you only fill in what you want to schedule:

• etas (SIGMAS) - a per-step curve for eta, the SDE noise-injection amount. This is the one most people actually use. Front-load it for noise early, taper it to zero for a clean finish.
• s_noises (SIGMAS) - a per-step curve for s_noise, the overall noise scaling. Pairs naturally with etas.
• unsample_resample_scales (SIGMAS) - per-step scaling for the unsample/resample workflow, which is RES4LYF's img2img-by-noise-inversion path.

The single output is automation (AUTOMATION), which goes to exactly one place: the sampler's automation input. Leave an input unconnected and that parameter just uses its normal fixed value.

Installing RES4LYF

Through ComfyUI Manager, search RES4LYF, install, restart. Or clone it:

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

Portable ComfyUI users point pip at the embedded interpreter under python_embedded/Scripts/. Restart and hard-refresh with F5. The pack downloads no models. rgthree-comfy is a recommended companion for the nested sampler menus.

Common issues

Nothing changed. If none of the three inputs is connected, this node is a no-op - it passes through and the sampler uses its plain widget values. You need at least one real SIGMAS curve plugged in.

The curve length doesn't match your steps. These are per-step schedules, so a curve built for 20 steps behaves oddly if you then render 30. Rebuild the curve to match, or keep the same step count you designed it around.

You're overcomplicating it. This is genuinely an advanced/experimental node - the pack's own docs flag much of this territory as "subject to further changes." If you're new to RES4LYF, get good results from a fixed eta first. Automation is the last 10%, not the starting point.