CFG Selector

AUNCFG is a single float slider that outputs a CFG (classifier-free guidance) scale value. That's it. One float input - range -2 to 100, step 0.1 - and one FLOAT output you wire into a KSampler's cfg socket. It's the "CFG Selector" in the node menu, and it exists so you can park the most-tweaked number in your workflow somewhere visible instead of burying it inside a collapsed sampler node.

Why would you bother when the KSampler already has a CFG field? Because CFG is the one value you'll re-tune every single time you change checkpoint, and on a busy canvas the sampler's widgets are often the first thing you fold away. Pull it out onto its own node and you get a bigger, cleaner slider you can label, plus a float output you can branch into other places - a filename token, a comparison setup, a second sampler that needs the same value. It's the same "extract the parameter you actually fiddle with" philosophy as the pack's KSamplerInputs node, just for the one number.

The finer control the description mentions is the step size: 0.1 instead of the coarser increments some samplers expose, which matters when you're hunting the edge between "under-inked" and "burned" at CFG 7.4 versus 7.5. The node's own tooltip tells you 7-12 is typical for many models, which is fair for SD 1.5 and SDXL. But here's the thing the knowledge base keeps hammering: CFG's sweet spot is architecture knowledge, not taste. Pony and Illustrious like 4-6, and anything guidance-distilled - Turbo, Lightning, most of 2026's fast checkpoints - runs at exactly 1, where ComfyUI skips the negative pass entirely for a real speed win. The tooltip's "7-12" is the SD 1.5-era default; trust the checkpoint's release notes over the generic guidance.

The negative range deserves a mention because it's rare: with min: -2 you can push CFG negative, which is a genuinely experimental thing some users do with certain samplers for stylized or inverted-guidance looks. Most people will never touch it, but it's there.

Mechanically the node is trivial - a widget in, the same float out - and that's exactly the point. No hidden transforms.

Install the pack via ComfyUI Manager (search "AUN") or cd ComfyUI/custom_nodes && git clone https://github.com/loz2754/AUN-ComfyUI-Nodes, then restart. Like the rest of the pack's utility nodes it's dependency-light (piexif, opencv-python-headless, imageio-ffmpeg, requests) and needs no model downloads. If you're on a distilled model and your images look overcooked, drop this node in, set 1, and watch the "prompt adherence" panic vanish.