Legacy_ClownSampler

RES4LYF went through a rewrite, and rather than break every workflow that used the old nodes, the pack kept the previous generation around with "Legacy" prepended to their names. Legacy_ClownSampler is one of those: the standalone SAMPLER-producing node from before the current all-in-one ClownsharKSampler. If you opened someone's older saved workflow and found this node instead of the modern one, that's exactly why - it still works, it's just not what you'd start a new workflow with today.

Where the current ClownsharKSampler bundles model, latent, guides, and scheduler into one node, this legacy version does one job: it outputs a SAMPLER object you plug into a regular ComfyUI KSampler (or KSampler Advanced), the same way you'd plug in any stock sampler.

How it works

This is RES4LYF's core solver engine - 69 samplers is a pack-wide figure, and this node exposes the pieces of that system directly. eta controls how much noise gets added after each SDE step; noise_mode sets how that amount scales (hard/hard_sq/soft/softer in increasing gentleness, or exp for values well above 1.0). noise_sampler_type picks the texture of that injected noise - fractal, gaussian, pyramid variants, and more. rk_type is the actual solver: names ending in "m" (multistep, like res_2m) reuse previous steps as estimates and run at Euler speed; names ending in "s" (substep, like res_2s) cost one model call per digit in the name but buy real accuracy, especially with noise. The res family are the pack's headline refinements of the dpmpp lineage. multistep and implicit_steps layer refinement passes on top (diminishing returns past 2-3), and the latent_guide*/guide_mode group adds optional img2img-style steering toward a reference latent.

The inputs and outputs that matter

• rk_type - the solver itself. res_2m is a solid, fast default; the s-suffixed variants (res_2s, res_3s) trade speed for accuracy.
• eta / noise_mode / noise_sampler_type - SDE noise strength, its scaling curve, and its texture. eta at 0 disables SDE injection entirely (pure ODE).
• noise_seed (default -1 increments the last seed used).
• latent_guide / latent_guide_inv / latent_guide_mask / latent_guide_weight / guide_mode - optional guidance toward a reference latent, useful for img2img-style work.
• sigmas_override - feed a custom sigma schedule here to bypass whatever scheduler your KSampler is using.
• Output: SAMPLER - wire this into a SamplerCustom-style node's sampler slot, the same place any stock sampler object would go.

How to install it

Via ComfyUI Manager: search "RES4LYF" and install. Manually, from ComfyUI/custom_nodes:

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

Activate your venv first if you use one; on portable installs, use python_embedded\Scripts\pip.exe in place of pip. Restart ComfyUI afterward. If you also want the newer nested sampler-picker menu (not needed for this legacy node to function, but useful if you're comparing against the current one), install rgthree-comfy too and enable "Auto Nest Subdirectories" in its settings.

Common issues & troubleshooting

Should you actually be using this? Only if you're loading an old saved workflow that already has it. For anything new, the current ClownsharKSampler is the maintained, documented path - this node is a compatibility shim, not the recommended starting point.

Generation is much slower than expected. Check rk_type - anything ending in a high digit-s suffix (res_5s, for instance) costs that many model calls per step. That's the accuracy/speed trade-off working as intended, not a bug.

Results look flat or over-smoothed. Try a nonzero eta with noise_mode on soft or softer - pure ODE sampling (eta = 0) can look duller than SDE sampling on some solvers, and this pack's whole premise is that controlled noise injection helps.