RescaleCFGAdvanced

Your image looks plastic. Skin has that waxy, over-saturated sheen, highlights are blown out, and dropping the CFG fixes the burn but kills prompt adherence. The classic answer is the core ReforgeCFG node (you'll also see it called CFG Rescale), which rescales the CFG output so its magnitude matches the conditional prediction - details come back without dialing CFG down. The problem: run it for the whole denoise and you get weird glitches, because the rescale is only meant for part of the process. RescaleCFGAdvanced is the version that lets you say which part.

It sits in advanced/model and does one thing ComfyUI core's ReforgeCFG doesn't: start_percent and end_percent inputs that gate when the rescale fires. The author's own before/after shows the sweet spot is roughly 0.05 → 0.5 of the sampling progress - early structure-setting steps, not the final polishing ones. Used that way it's genuinely nice on models with plastic skin, which is exactly the case people reach for it on.

How it works

The trick behind CFG rescale comes from the paper Rescaling Classifier-Free Guidance. Plain CFG is a lerp from the uncond prediction toward the cond one, scaled by your CFG value. Problem is, that amplifies the whole latent uniformly, so areas that need less correction get blown out while others stay flat. Rescale measures the standard deviation of the conditional prediction against the standard deviation of the CFG output, and shrinks the output's magnitude to match - pushing detail back into over-baked regions without touching prompt adherence.

Under the hood this node clones your model and patches its sampler CFG function. Each step it reads the sigma schedule, works out what percentage through the run you are, and only applies the rescale math when you're inside [start_percent, end_percent]. Outside the window it computes plain uncond + cond_scale * (cond - uncond) and moves on. It internally converts to a v-prediction parameterization to do the rescale and back, which is why it behaves well on v-pred models - that's the other spot you'll see it recommended, on Illustrious vpred setups alongside CFG++ samplers.

The inputs that matter

Three of them, plus the model. Keep an eye on:

• model - your checkpoint output (or a patched chain of it). Wire the output into your sampler's model slot.
• start_percent / end_percent - the sampling-progress window where rescaling is active. The README's demo image is the whole argument: run it 0 → 1 and you get glitches; 0.05 → 0.5 looks good. If you accidentally set start above end, the node swaps them for you - no crash, just don't rely on it.
• multiplier (default 0.7) - blends between normal CFG and full rescale. 1.0 is pure rescale, 0.0 is plain CFG, and the in-between is where you'll actually live once you trust the node.

Output is a single MODEL, which you feed to your KSampler. That's it - a drop-in patch.

Installing it

ComfyUI Manager (search "RescaleCFGAdvanced"), or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/BigStationW/ComfyUi-RescaleCFGAdvanced

Restart ComfyUI, double-click the canvas and search for it. No extra dependencies, no model files, nothing to download - the whole node is one self-contained Python file. Nice change of pace.

Common gotchas

The defaults (start 0, end 1) are the worst setting for the reason you installed it - full-range rescale is what causes the glitches. Change them before you judge the node.

Second, the percentage gating needs the sampler's sigma schedule to be visible to the patch. If it can't find it, the node quietly falls back to standard CFG unless your range is the full 0 → 1. In practice this works with the normal KSamplers, but if you're using an exotic sampler setup and the rescale silently isn't happening, that's the first thing to suspect.

Don't stack it on a guidance-distilled model that's already running CFG 1 - there's no burn to fix when guidance is baked in, and the double-correction problem the rescale solves doesn't apply. This is a tool for real CFG models running in the 5–9 range. Start at the author's 0.05 → 0.5 with multiplier 0.7, then widen the window only if you're still seeing burned highlights.