Mepi Steps & CFG

Two dials, passed through, that's it

Mepi Steps & CFG takes a steps number and a cfg number and returns them as STEPS and CFG outputs, ready to plug into a KSampler. The source is return (steps, cfg,) - the author even labeled it a "Dirty copy of Mira's StepsAndCfg," so you're looking at a community utility that got repackaged for ComfyMepi, a small mobile frontend for ComfyUI by rhplus0831. The phone UI can't fiddle with sampler inputs directly, so these Mepi nodes are the knobs it writes into before resubmitting the workflow.

What you set:

• steps (INT, default 30, min 1) - sampling steps.
• cfg (FLOAT, default 7.0, min 1) - classifier-free guidance scale.
• STEPS and CFG outputs - wire into the KSampler's steps and cfg inputs.

The defaults are the classic SD 1.5-era starting point: 30 steps and CFG 7. That combination has been a reasonable baseline for years, and it's still fine for SD-lineage models.

What CFG actually is (one paragraph, since you're staring at it)

CFG is how hard the model is pushed toward your prompt at every denoising step - it predicts each step twice, once with your prompt and once without, and amplifies the difference. Too low and the model ignores you; too high and images come out saturated and "deep fried." But here's the thing that's quietly eaten this dial: the 2026-generation models don't want 7 at all. Guidance-distilled and flow-matching models (Flux and friends) run at CFG 1 by design - and at exactly 1, ComfyUI skips the unconditional pass entirely, so it's also faster. If you're switching model families and your images look off, reach for this node's cfg before anything else.

Install

Nothing special, same as the whole pack. Via ComfyUI Manager: Install Custom Nodes → search "ComfyMepi", or:

cd ComfyUI/custom_nodes
git clone https://github.com/rhplus0831/ComfyMepi

Restart ComfyUI. No requirements.txt, no pip install, no model downloads.

Troubleshooting

• Overcooked, burned, high-contrast images: classic too-many-steps-plus-too-high-CFG. The community's standard fix is drop to ~30 steps and CFG 7, then lower CFG if it's still harsh. Since this node defaults to exactly that, it's a fine place to start.
• Washed-out, prompt-ignoring images: CFG too low for the model, or you're on a distilled model where the rules are just different. On SD-lineage, raise CFG; on Flux-style models, the answer is the opposite of what your muscle memory says.
• Why can't I set cfg below 1? The minimum here is 1.0. Some niche setups want sub-1 guidance, but for distilled models 1 is exactly what you want, so it rarely bites.

The pack is a personal project with essentially no community footprint and light maintenance - worth knowing, since this node is where you'd file a feature request for a higher step range. But as a pass-through with sensible defaults, it's about the least likely node in the pack to cause you grief.