All Parameters Loop

This is the node the pack's name promises and the others only gesture at. All Parameters Loop cycles through everything at once - steps, CFG, shift, sampler, and scheduler - and hands each combination to your KSampler one run at a time. The example workflow bundled with the pack shows the intended shape: its five value outputs plug straight into a KSampler's steps, cfg, sampler, and scheduler inputs, and the loop node's current_combination string goes to a text preview so you can see what each run is doing.

If you're tuning a new model - especially a WanVideo setup, where this node's shift output actually lands in the WanVideo sampler - it's a genuine timesaver. Set your ranges, queue N times, and get a labeled sweep instead of hand-editing five widgets between every run.

The mechanism, briefly

One internal counter per mode, and the node advances it only when it sees a new execution (it dedupes via a timestamp-plus-thread ID). So: one queue run = one combination, next run = next combination. Same rule as every node in this family - "loop" means "roll forward on each execution," not "iterate internally."

Combinations are enumerated in a fixed order: steps, then CFG, then shift, then sampler, then scheduler. So the last range you set changes fastest. Sequential mode walks that product in order; random mode picks an index using seed + step; ping_pong bounces back and forth through the whole product.

Inputs that matter

The three ranges use _start / _end / _interval naming, which is just the pack's way of saying "from, to, by":

• steps_start / steps_end / steps_interval - integers, defaults 20/50/10. Steps is always the integer slider.
• cfg_start / cfg_end / cfg_interval - floats, defaults 1/8/1. CFG behavior varies hard by model family - Wan, for instance, wants CFG 1.0 with speed LoRAs and 3.5+ without (the KB's Wan knowledge doc spells this out), so range-sweeping CFG is where this node earns its keep.
• shift_start / shift_end / shift_interval - floats, defaults 1/3/0.5. shift is the flow/sigma shift parameter the WanVideo samplers use (lightx2v workflows typically run shift ~8). On SD/SDXL graphs it has no matching input - the pack's own example just leaves it unplugged.

Plus the usual mode, seed, reset, and the skip_samplers / skip_schedulers comma lists. Floats are rounded to two decimals internally to dodge floating-point drift, which is a nice touch you don't notice until you'd otherwise be chasing off-by-one bugs.

Outputs: steps, cfg, shift, sampler (a string, stock-KSampler-compatible), scheduler (a string), then current_index, total_combinations, and current_combination.

Do the math before you queue

total_combinations is the product of all five ranges. Three step values × 8 CFG values × 5 shift values × 44 samplers × 9 schedulers = 47,520 runs. On video scales that's a month of GPU. This node is a trap if you leave the defaults wide and just hit queue.

Trim with the skip lists and narrow ranges. My move: fix sampler and scheduler first (skip everything but the two or three you believe in), sweep steps/CFG/shift, then lock those and let the sampler loop go. The reset boolean (BOOLEAN, default false) zeroes the counter whenever you want to restart a clean sweep.

Install

No dependencies, no models - this whole pack is pure Python over ComfyUI's own sampler/scheduler lists. ComfyUI Manager, search comfyui-ksampler-tester-loop (or "Sampler Scheduler Loop"), install, restart. Manual route:

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

then restart ComfyUI. Check the console if something's off - the node prints warnings for bad skip-list names and logs every selection, which is surprisingly useful when you're staring at run #200 and can't remember what it's testing.

One honest caveat: with all those dials, a genuinely blind grid search is a bad idea - but as a structured sweep you control, this is the most useful node in the pack.