🐟Sampler Settings

If your workflow has more than one KSampler - a base pass and a hires-fix pass, say, or a couple of sampler variants you flip between - you've probably felt the annoyance of setting sampler_name, scheduler, steps, and denoise separately on each one, then forgetting to update all of them when you change your mind. SamplerSettings exists purely to fix that: set the four values once, wire the outputs to every sampler that needs them, and change your mind in exactly one place.

How it works

There's no sampling logic in this node at all - it's a settings bundle. It exposes the same sampler and scheduler dropdowns ComfyUI's own KSampler uses, plus steps and denoise, and just passes them straight through as outputs. The value is entirely in not having to duplicate four widget settings across every sampler node in a complex graph.

The inputs and outputs that matter

• sampler_name - all of ComfyUI's built-in samplers (Euler, Euler Ancestral, DPM++ 2M, and around 40 more).
• scheduler_name - the 9 stock ComfyUI schedulers: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, normal, linear_quadratic, kl_optimal.
• steps (INT, 1–100, default 30) and denoise (FLOAT, 0–1, default 1) - standard sampler knobs.

Outputs mirror the inputs - sampler_name, scheduler_name, steps, denoise - for wiring into one or more KSampler-style nodes. There's also a fifth output, summer (STRING), that isn't documented anywhere in the README and doesn't map to anything obvious from the node's stated purpose. Worth knowing it's there so you're not confused when you see an extra unlabeled socket, but you can safely leave it unwired.

How to install it

Search ComfyUI-QHNodes in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/liuqianhonga/ComfyUI-QHNodes.git

Restart. Main-repo node, no submodules, no downloads - pure configuration passthrough.

Common issues & troubleshooting

Karras by default, and it's the wrong call on newer models. If you're on a flow-matching checkpoint - Flux, Z-Image, Anima, or basically anything released since 2024 that isn't an SDXL finetune - Karras is not a safe default the way it was for SD 1.5/SDXL. Karras concentrates denoising effort mid-trajectory, which corrects for a curved DDPM sampling path; flow-matching models sample along a near-straight path, so that same correction becomes a distortion instead. Community sweeps on Z-Image Turbo specifically flag Karras and Exponential as consistent underperformers there. Check what architecture your checkpoint actually is before reaching for Karras out of habit.

Your scheduler list is missing something a workflow calls for. The 9 schedulers here are the stock ComfyUI set. Anything like beta57 or other RES4LYF-specific schedules won't show up - those live in a separate node pack (RES4LYF, by ClownsharkBatwing) with its own sampler/scheduler selector, not in this node.

Distilled or speed-LoRA models look burned or oversaturated. That's usually a steps/CFG mismatch rather than anything this node controls directly - step-distilled checkpoints (Lightning, Turbo, DMD2-style LoRAs) expect a step count matching their distillation (often 4-8), not the 20-30 this node defaults to. Match the steps to what the checkpoint or LoRA card actually recommends.