ComfyUI-CFG-Ctrl

SMC-CFG — Sliding Mode Control Classifier-Free Guidance for ComfyUI
Based on CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance (CVPR 2026)

---

What is it?

CFG-Ctrl reinterprets Classifier-Free Guidance (CFG) through the lens of control theory:

| Method | Description | Formula | |---|---|---| | Vanilla CFG | Proportional controller (P-control) with fixed gain | v = v_u + w * e | | SMC-CFG | Nonlinear sliding mode controller | v = v_u + w * e + u_sw |

Where:

• e_t = v_cond - v_uncond — guidance error (discrepancy between conditioned and unconditioned predictions)
• s_t = e_t + (λ - 1) * e_{t-1} — exponential sliding surface
• u_sw = -K * sign(s_t) — nonlinear switching correction

Why it matters: standard CFG at high scales causes instability and semantic overshooting. SMC-CFG adds a correction term that enforces convergence along a sliding manifold, preventing those failures with Lyapunov-guaranteed finite-time stability.

Installation

Clone this repo into your ComfyUI custom_nodes folder:

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

No extra dependencies required — uses only torch, which ComfyUI already ships with.

Nodes

SMC-CFG (CFG-Ctrl) — Simple

Plug-and-play node. Connect between the model loader and the KSampler.

| Input | Type | Default | Description | |---|---|---|---| | model | MODEL | — | Any ComfyUI model | | smc_lambda | float | 5.0 | Decay rate of the sliding surface (λ) | | smc_k | float | 0.2 | Switching gain K | | warmup_steps | int | 2 | Initial steps using vanilla CFG before SMC activates |

SMC-CFG Advanced (CFG-Ctrl) — Advanced

All parameters from the simple node, plus:

| Input | Type | Default | Description | |---|---|---|---| | switching_mode | enum | hard | hard = sign(s) (paper default) · smooth = tanh(s/δ) | | tanh_delta | float | 0.1 | Smoothing bandwidth δ (only for smooth mode) | | normalize_error | bool | False | Normalise the sliding surface to unit norm before applying K (makes behaviour more consistent across models) | | smc_end_step | int | 0 | Deactivate SMC after this many steps (0 = always active) |

Workflow

[Load Checkpoint] ──► [SMC-CFG (CFG-Ctrl)] ──► [KSampler]
                                                     ▲
                         [CLIP Text Encode] ─────────┤ (positive)
                         [CLIP Text Encode] ─────────┤ (negative)
                         [VAE] ──────────────────────┘ (latent)

The CFG scale on the KSampler is the proportional gain w. SMC adds the nonlinear correction on top — it does not replace your CFG setting.

Recommended Settings

| Model | CFG Scale | λ (lambda) | K | Notes | |---|---|---|---|---| | FLUX.1-dev | 2–3 | 5.0 | 0.2 | Use warmup_steps = 2 | | SD3 / SD3.5 | 7.5 | 5.0 | 0.2 | | | Qwen-Image | 4.0 | 5.0 | 0.2 | | | Wan Video | 5.0 | 5.0 | 0.2 | |

Note: Models that run at CFG = 1 (fully distilled, e.g. some Flux variants) will see no effect, since e_t = 0 in those cases.

How the State Reset Works

The node tracks e_prev (error from the previous denoising step) across steps within the same generation. It automatically resets at the start of each new generation by detecting:

1. Sigma increase — the scheduler jumped back to high noise (new run started).
2. Shape change — the latent tensor dimensions changed (different batch/resolution).

Troubleshooting

| Symptom | Fix | |---|---| | Images look the same as without the node | Make sure CFG scale > 1 on KSampler | | Artifacts / over-saturation | Lower smc_k (try 0.05–0.1) | | Output too similar to negative prompt | Raise smc_lambda | | Chattering / noisy edges | Switch to smooth mode and set tanh_delta = 0.2 |

Citation

@misc{wang2026cfgctrlcontrolbasedclassifierfreediffusion,
  title   = {CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance},
  author  = {Hanyang Wang and Yiyang Liu and Jiawei Chi and Fangfu Liu and Ran Xue and Yueqi Duan},
  year    = {2026},
  eprint  = {2603.03281},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2603.03281},
}

License

MIT — do whatever you want, credit appreciated.