SMC-CFG Advanced (CFG-Ctrl)

The plain SMC-CFG node is plug-and-play, and that's its whole appeal. The Advanced version is what you graduate to when the simple one is almost right - edges chattering with noise, the correction running too hot late in the denoising run, or just wanting the paper's exact formulation instead of the safe defaults. Same model-patch trick, same sliding-mode math, but every knob is on the table.

How it's different from the simple node

Both nodes clone the model and register a custom sampler_cfg_function that recomputes guidance as v_uncond + w * (e_t + u_sw), where u_sw = -K * φ(s) is the sliding-mode correction from CFG-Ctrl (arXiv 2603.03281). The Advanced node just stops hiding the internals. The simple node hardcodes "hard" switching and one smoothing bandwidth; here you get to shape the correction term itself.

The extra inputs

• switching_mode (default smooth) - hard = sign(s), the paper's discontinuous default; smooth = tanh(s/δ), a differentiable approximation that trades a little sharpness for far less chattering. This is the main lever for noisy edges.
• tanh_delta (default 0.5) - the smoothing bandwidth δ. Smaller = closer to a hard sign. If smooth mode feels mushy, drop it toward 0.2.
• smc_end_step (default 0) - deactivates SMC after N steps and falls back to vanilla CFG. 0 = always active. This is the one I'd reach for when the last few steps are wrecking an otherwise good render, since CFG overshoot damage concentrates at the end of a run.
• normalize_s (default off) - a legacy option that normalizes the sliding surface to unit norm before the switching function, independent of adaptive_scale. The tooltip says it plainly: adaptive_scale is the preferred cross-model safety net. Leave this alone unless you're chasing paper parity.

You also get the simple node's whole input set - model, smc_lambda, smc_k, warmup_steps, adaptive_scale, max_correction_ratio - with one notable difference: smc_k ranges up to 2.0 here instead of 1.0. The Advanced node assumes you trust the clamp (max_correction_ratio) to keep the correction from running away, so it gives K more room.

Output is a single model (MODEL), wired into the KSampler. Nothing else changes in your graph.

Install

Same pack as the simple node - one install covers both. In ComfyUI Manager, search "CFG-Ctrl", install, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/flyghtxmz/ComfyUI-CFG-Ctrl

No requirements.txt, no model files, only torch. Restart and both nodes show up under the CFG-Ctrl category.

Troubleshooting with the new knobs

• Chattering / noisy edges - switch to smooth mode and set tanh_delta around 0.2. That's the README's own fix, and it's usually the first thing people hit with the paper's hard sign.
• Correction blows up or ruins late steps - either trust max_correction_ratio more (it's a hard clamp on ||u_sw|| as a fraction of ||e_t||) or gate SMC off entirely with smc_end_step.
• Not sure which knob is hurting you - this node, unlike the simple one, lets you burn a lot of time turning dials. The defaults (adaptive_scale on, smooth mode, λ 0.05, K 0.1) are the safe harbor; the paper's λ is 5.0 if you want the full-strength behavior.
• No effect at all - CFG 1 or a distilled model. There's no cond/uncond error signal to correct, and no switching term can conjure one.

One thing worth knowing before you over-tune: this is a very young pack (the CFG-Ctrl paper only dropped March 2026, the repo days later), so there's no deep community bench of settings for it. The one real integration I've seen is a Wan I2V workflow that runs it alongside NAG for prompt-following. And the standard skeptic's caveat applies double here: if you're comparing modes, change one thing at a time on a fixed seed, and judge across several generations - a single lucky render proves nothing.