RES4LYF Sampler Loop

The problem this node solves is the one everyone hits the moment they install RES4LYF: you now have well over a hundred samplers, and none of them are in ComfyUI's stock dropdown. The honest way to pick between them is to run the same prompt and seed through a bunch and look at the results side by side. That's the "change → queue → rename → repeat" grind, and RES4LYFSamplerLoop is a small machine built to automate exactly that.

It's the flagship node of KY-2000's RES4LYF-tester-loop pack, a companion to ClownsharkBatwing/RES4LYF - the sampler pack that, for flow-matching models like Flux, Z-Image and Anima, has basically replaced DPM++ 2M Karras as the place sampler tuning happens. The loop nodes don't add samplers; they let you iterate over the ones you already have, one per queue run, so every image in a batch comes out with a different solver. The author's own sample workflow wires this straight into RES4LYF's ClownsharKSampler_Beta node, with AddLabel captioning each output.

How it works

On every execution the node advances a counter and picks the next sampler from your actual RES4LYF install - it reads the enum list directly from RES4LYF/beta/rk_coefficients_beta.py at import time, so you get exactly the solvers you really have, including new ones added upstream. If that file isn't found (RES4LYF missing or installed under a different folder name), it falls back to a built-in list of 100+ samplers and prints a [RES4LYF Tester Loop] Using FALLBACK_SAMPLERS line in the console, which is your cue that something's off.

Three modes control selection: sequential walks the list in order, random picks via a seeded RNG, and ping_pong bounces back and forth through the range. seed makes random reproducible; reset (true) pushes the counter back to the start.

The inputs that matter

• mode - sequential / random / ping_pong. The one you'll actually change.
• seed - only relevant in random mode.
• reset - flip it once to restart the sweep.
• skip_samplers - comma- or newline-separated names to exclude, e.g. none, exponential/res_16s.
• loop_sampler_mode + sampler_mode_method + skip_sampler_modes - optionally cycle RES4LYF's three sampler modes (unsample, standard, resample) on top.

Outputs: sampler_name (the enum, wired into ClownsharKSampler_Beta.sampler_name), current_index, total_combinations, current_combination, sampler_mode, and a label like multistep/res_2m | standard you can push into AddLabel or a filename.

Install

Two things, in order: RES4LYF first, then this pack, both in ComfyUI/custom_nodes/.

cd ComfyUI/custom_nodes/
git clone https://github.com/ClownsharkBatwing/RES4LYF.git
git clone https://github.com/KY-2000/RES4LYF-tester-loop.git

Restart ComfyUI. ComfyUI Manager also finds both by title. No pip dependencies, no model downloads - the whole pack is stdlib Python.

Where people get burned

The sampler enum literally starts with "none", a real option that does nothing useful in a sweep - add it to skip_samplers. All instances of this node class share one counter, so two Sampler Loops in one workflow advance together; use one per sweep or you'll be confused. And remember the sweep will happily hand you Karras/exponential pairings that are known to fail on flow-matching models - that's fine, it's a sweep, just don't mistake a dead combo for a verdict on the sampler itself.