Initial implementation of SMC-CFG Ctrl ComfyUI node

Implements the Sliding Mode Control CFG algorithm from the paper
"CFG-Ctrl: A Control-Theoretic Perspective on Classifier-Free Guidance" (CVPR 2026)
as a ComfyUI model patch node.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

__init__.py

@@ -0,0 +1,3 @@
+from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]

nodes.py

@@ -0,0 +1,102 @@
+import torch
+
+
+class SMCCFGCtrl:
+    """
+    Implements SMC-CFG (Sliding Mode Control CFG) from the paper:
+    "CFG-Ctrl: A Control-Theoretic Perspective on Classifier-Free Guidance" (CVPR 2026)
+    https://github.com/hanyang-21/CFG-Ctrl
+
+    Replaces standard linear CFG with a nonlinear sliding mode controller
+    that prevents instability, overshooting, and artifacts at high guidance scales.
+    """
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "model": ("MODEL",),
+                "smc_cfg_lambda": ("FLOAT", {
+                    "default": 5.0, "min": 0.0, "max": 50.0, "step": 0.01,
+                    "tooltip": "Sliding surface coefficient. Controls how much the controller weights previous error magnitude vs error derivative. Paper recommended: 5.0",
+                }),
+                "smc_cfg_K": ("FLOAT", {
+                    "default": 0.2, "min": 0.0, "max": 5.0, "step": 0.01,
+                    "tooltip": "Switching gain. Bounds the correction to [-K, +K] per element. Higher = stronger correction but may introduce chattering. Paper recommended: 0.2",
+                }),
+                "warmup_steps": ("INT", {
+                    "default": 0, "min": 0, "max": 100,
+                    "tooltip": "Number of initial steps with no guidance (pure conditional prediction). Lets the model establish structure before guidance kicks in.",
+                }),
+            }
+        }
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+    CATEGORY = "sampling/custom_sampling"
+
+    def patch(self, model, smc_cfg_lambda, smc_cfg_K, warmup_steps):
+        # Mutable state persisted across denoising steps via closure
+        state = {
+            "prev_eps": None,
+            "step": 0,
+            "prev_sigma": None,
+        }
+
+        lam = smc_cfg_lambda
+        K = smc_cfg_K
+
+        def smc_cfg_function(args):
+            cond = args["cond"]
+            uncond = args["uncond"]
+            cond_scale = args["cond_scale"]
+            sigma = args["sigma"]
+
+            # Detect new generation: sigma should decrease monotonically during
+            # denoising. If it jumps up, a new sampling run has started.
+            curr_sigma = sigma.max().item() if torch.is_tensor(sigma) else float(sigma)
+            if state["prev_sigma"] is not None and curr_sigma > state["prev_sigma"] * 1.1:
+                state["prev_eps"] = None
+                state["step"] = 0
+            state["prev_sigma"] = curr_sigma
+
+            step = state["step"]
+            state["step"] = step + 1
+
+            # Warmup: pure conditional prediction (no guidance)
+            if warmup_steps > 0 and step < warmup_steps:
+                return cond
+
+            # Guidance error: e_t = noise_cond - noise_uncond
+            guidance_eps = cond - uncond
+
+            if state["prev_eps"] is not None:
+                prev_eps = state["prev_eps"]
+
+                # Sliding surface: s_t = (e_t - e_{t-1}) + lambda * e_{t-1}
+                s = (guidance_eps - prev_eps) + lam * prev_eps
+
+                # Switching control: u_sw = -K * sign(s_t)
+                u_sw = -K * torch.sign(s)
+
+                # Apply correction to guidance error
+                guidance_eps = guidance_eps + u_sw
+
+            # Store corrected guidance for next step's sliding surface
+            state["prev_eps"] = guidance_eps.detach().clone()
+
+            # v_guided = v_uncond + scale * corrected_guidance
+            return uncond + cond_scale * guidance_eps
+
+        m = model.clone()
+        m.set_model_sampler_cfg_function(smc_cfg_function, disable_cfg1_optimization=True)
+        return (m,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "SMCCFGCtrl": SMCCFGCtrl,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "SMCCFGCtrl": "SMC-CFG Ctrl",
+}