Full K corrupts at high CFG (output correction = cfg*K = 2.4 at cfg=12).
K/cfg was too weak (0.2 at cfg=12). The paper only tested up to cfg=7.5
where output corrections range 0.5-1.5.
K/sqrt(cfg) keeps output correction = sqrt(cfg)*K growing sub-linearly,
giving 0.69 at cfg=12 — within the paper's working range.
Also store raw (pre-correction) guidance as prev_eps to prevent
correction accumulation through the sliding surface.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The paper stores corrected guidance, but in ComfyUI the corrections
compound through the sliding surface's lambda * prev_eps term
(amplified 4x per step at lambda=5). Over 20 steps this overwhelms
the actual guidance signal, causing total corruption at full K.
Storing raw (pre-correction) guidance keeps the surface tracking
the model's actual guidance evolution while applying fresh
corrections each step. This allows using full K=0.2 (matching
the paper) without accumulation-driven instability.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The K_eff = K/cond_scale compensation was making the correction
too weak at high CFG (0.017 at cfg=12 vs paper's 0.2). The original
paper uses full K and relies on cfg_scale * K amplification for
stabilization. tanh smoothing + 5x5 spatial blur handle artifact
prevention that the paper doesn't need in DiffSynth.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The per-element correction creates a visible mesh pattern at the VAE's
8x8 patch boundaries. A 3x3 box blur in latent space (24x24 pixels)
smooths adjacent corrections while preserving the large-scale
correction structure.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two remaining noise sources fixed:
1. CFG amplification: the return value is uncond + cond_scale * u_sw * sigma,
so the noise-space correction is cond_scale * K per element. At cfg=12
with K=0.2, that's 2.4 — far too large. Fix: K_eff = K / cond_scale,
making the effective correction just K regardless of cfg scale.
2. Hard switching: even clamp(s/phi) creates sharp transitions at the
boundary. Replace with tanh(s/phi) for a fully smooth correction.
phi = std(s) normalizes the sliding surface to its natural scale.
Net effect: the noise-space correction is now bounded by K=0.2 per element
regardless of cfg scale, and varies smoothly across spatial positions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The hard sign(s) creates a random +/-1 pattern in regions where the
guidance error is near zero. When amplified by cond_scale (e.g. 12),
this produces severe high-frequency noise artifacts, especially on
SDXL which has smaller guidance magnitudes than SD3/FLUX.
Replace sign(s) with sat(s/phi) — the standard boundary layer approach
in practical sliding mode control. phi adapts to the guidance error
std so the switching threshold is meaningful across different models.
- |s| >> phi: correction = +/-K (same as paper)
- |s| << phi: correction = K*s/phi (smooth, proportional)
- Near-zero guidance regions get near-zero correction (no noise)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous fix (denoised space) still had the problem: K * cond_scale
produced a constant ±2.4 perturbation per element at cfg=12, destroying
the image at every step.
The paper's K=0.2 is calibrated for unit-variance noise predictions.
ComfyUI's cond/uncond are sigma-scaled (x - denoised ≈ sigma * epsilon).
Now we divide by sigma to recover epsilon-space, apply SMC there, then
multiply back by sigma. This gives natural dampening at late steps:
- sigma=14 (early): correction ±33 in latent space (image is noise anyway)
- sigma=0.01 (late): correction ±0.024 in latent space (negligible)
This matches the paper's behavior where the scheduler conversion
inherently dampens the noise-space correction at low sigma values.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ComfyUI's args["cond"]/["uncond"] are (x - denoised), which are
sigma-scaled. At late denoising steps (sigma~0.01), the fixed K=0.2
correction was 200x the signal magnitude, destroying the image.
Fix: compute SMC in denoised space using args["cond_denoised"] and
args["uncond_denoised"], which have consistent magnitude across all
sigma values — matching the paper's noise-prediction space.
Also fixes first-step behavior to match the original paper (SMC
correction applied from step 0, not step 1).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
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>
