Skimmed CFG - linear interpolation

If there's one node from the Skimmed_CFG pack that a lot of people end up keeping in their daily workflow, it's this one: the linear-interpolation skim. It does the same anti-burn job as the base node - cap how far the CFG extrapolation can push the latent at each step - but instead of snapping the conflicting values to a hard fallback scale, it blends between the positive and negative predictions where they fight. The transition is soft, and soft is what makes it look cleaner in practice. The author himself flags the linear interpolation variants as "highly recommended."

The typical recipe, from people who've written it up: set the sampler CFG high (12–15 or more), put 2–5 in the skimming field, and get the prompt adherence of the high value without the oversaturation. It's a known fix for the "colors go neon at CFG 4+" problem too.

How it works

The node registers a pre-CFG function on a cloned model. Each step it builds the same "outer influence" mask as the base node - latent elements where the positive and negative predictions pull against each other hard enough that the full CFG would overshoot. At those positions it linearly interpolates the uncond toward the cond, with a weight computed from your skimming scale relative to the sampler's CFG ((skimming − 1) / (sampler CFG − 1)). Everything else keeps full CFG. The result reads as "high adherence where it's safe, gentle capping where it would burn."

Wire it model loader → node → KSampler. Output is a MODEL.

The inputs

• model - MODEL in.
• skimming_cfg (default 5, 0–10) - the fallback scale the interpolation leans toward. Lower is more aggressive anti-burn; 5–7 trends toward stronger color. One number, that's the whole interface.

Output: MODEL.

Install

Part of Extraltodeus/Skimmed_CFG - single Python file, zero pip deps, no model downloads. ComfyUI Manager: search "Skimmed_CFG" (Extraltodeus) → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Extraltodeus/Skimmed_CFG

Restart after either.

Troubleshooting

• Inert at CFG 1. No unconditioned pass, no skim. On guidance-distilled models (Flux, turbo checkpoints) you must raise the sampler CFG first - and at that point this is precisely the tool that keeps it from burning.
• Too low a skimming scale means more steps; extremely low skim + very few steps can fuse fingers.
• SDE samplers can still burn a bit at big scales; the Timed flip node is the SDE-friendly sibling.
• This is the current name for what used to be "Skimmed CFG - linear interpolation." Old workflows using that name can complain after the January 2026 rename - this node is the drop-in.