ComfyUI-SMC-CFG
ComfyUI wrapper for SMC-CFG (Sliding Mode Control) classifier-free diffusion guidance from CVPR 2026 research.
ComfyUI-SMC-CFG (Official Bridge)
This repository provides a ComfyUI for desktop wrapper for SMC-CFG (Sliding Mode Control), as introduced in the CVPR 2026 paper:
"CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance", https://github.com/hanyang-21/CFG-Ctrl.
📖 Introduction
CFG-Ctrl reinterprets Classifier-Free Guidance as a control problem. While vanilla CFG acts as a proportional controller (P-control), it can lead to instability at high scales. SMC-CFG introduces a nonlinear sliding mode controller that enforces the guidance error to converge along an exponential sliding surface, preventing "overbaking" and improving semantic alignment.
🛠 Installation
- Navigate to
ComfyUI/custom_nodes. - Clone this bridge:
git clone https://github.com/belgiangurista-art/ComfyUI-SMC-CFG. - Install the base requirements:
pip install diffsynth. - Ensure you have the original research code available if running custom scripts.
📊 Official Parameters & Recommended Settings
Based on the original repository:
Note: For distilled/lightning models (like 4-step variants), doesn't work, results are overbaked.
🎓 Citation & Credits
All credit goes to the original authors: Hanyang Wang, Yiyang Liu, Jiawei Chi, Fangfu Liu, Ran Xue, Yueqi Duan (Tsinghua University).
Original Source: hanyang-21/CFG-Ctrl.