Inference_Core_RescaleClassifierFreeGuidanceTest

Like its neighbor in the experiments bucket, this node's class name is honest about its status - "Test" - even though the technique behind it is a real, published fix for a well-known problem: pushing CFG high enough for strong prompt adherence tends to burn the image, showing up as oversaturated colors, blown-out contrast, and deformed faces, roughly above CFG 12 on SD-lineage models and lower still on more sensitive checkpoints. CFG rescaling is the proposed fix - instead of using the CFG-guided prediction as-is, you rescale its statistics back toward the unguided prediction's by some amount, keeping the adherence gain from high CFG while pulling the exposure and contrast back from the edge before it collapses into artifacts.

The one knob

multiplier (0–1, default 0.7) is the rescale strength. At 0, it's effectively disabled - behaves exactly like plain CFG, no rescaling at all. At 1, it rescales fully back to the unguided prediction's statistics, which tends to look flat and washed out, since you're throwing away most of what CFG was actually doing for you. The useful range sits in between, and the shipped default of 0.7 is a reasonable starting point to tune from, not a value to treat as gospel - different checkpoints and different CFG values will want slightly different amounts of rescaling.

model in, patched MODEL out. Wire it into your KSampler like any other model patch, and keep your CFG wherever you actually want it for prompt adherence - this node doesn't replace CFG tuning, it changes what happens once CFG is already pushed high.

When it's actually worth adding

Not a default addition to every workflow. Most checkpoints look fine within their normal CFG range and gain nothing from this - SD 1.5 and SDXL's sweet spot is roughly 5-9, Pony and Illustrious sit lower at 4-6, and anything guidance-distilled (Flux, Turbo/Lightning checkpoints) runs at CFG 1 where this node has nothing to correct in the first place. Reach for it specifically when you deliberately want to run CFG higher than the normal range - for stronger, more literal prompt adherence - and want to avoid paying the usual burn tax for doing so.

Installing it

Bundled with the pack. ComfyUI Manager: search "ComfyUI-Inference-Core-Nodes". Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes

then install.py, or pip install -e . and restart - no acceleration extras needed, it's a sampler-level patch rather than a separate inference model.

Troubleshooting

If output looks unexpectedly flat or desaturated after adding this node, multiplier is almost certainly set too high - back it down toward 0 and re-test. If you don't see any difference at all, you're probably already running CFG low enough that there was no burn to correct in the first place; this node has nothing to do at CFG 1, and not much to do inside a checkpoint's normal recommended range. Because it's flagged as experimental, don't assume it plays identically across every checkpoint family - verify on your actual model before relying on it in a production workflow.