Ethanfel/ComfyUI-CFG-CTRL
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix correction accumulation: store raw guidance in prev_eps

Browse Source 
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>
This commit is contained in:
Ethanfel
2026-03-04 18:59:39 +01:00
parent 3c92369305
commit 2c50da025d
1 changed files with 11 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
nodes.py
View File

@@ -90,15 +90,8 @@ class SMCCFGCtrl:
# Smooth switching via tanh(s/phi) instead of hard sign(s). # Smooth switching via tanh(s/phi) instead of hard sign(s).
# The paper uses sign(s) which works in DiffSynth but creates # The paper uses sign(s) which works in DiffSynth but creates
# salt-and-pepper artifacts in ComfyUI's latent space. tanh # salt-and-pepper artifacts in ComfyUI's latent space. tanh
# provides the same bounded correction with smooth spatial # provides smooth spatial gradients: proportional near zero,
# gradients: proportional near zero, saturating at ±K for # saturating at ±K for large |s|.
# large |s|. phi normalizes s so the transition width matches
# the actual surface magnitude distribution.
#
# K is used at full strength (matching the paper) — the
# tanh smoothing + spatial blur handle artifact prevention.
# The paper's correction of cfg_scale * K per element in noise
# space is what provides the stabilization at high CFG.
phi = s.std().clamp(min=1e-6) phi = s.std().clamp(min=1e-6)
u_sw = -K * torch.tanh(s / phi) u_sw = -K * torch.tanh(s / phi)
@@ -107,14 +100,17 @@ class SMCCFGCtrl:
if u_sw.ndim == 4: if u_sw.ndim == 4:
u_sw = F.avg_pool2d(u_sw, kernel_size=5, stride=1, padding=2) u_sw = F.avg_pool2d(u_sw, kernel_size=5, stride=1, padding=2)
# Corrected guidance error (in normalized noise space) # Store RAW guidance (before correction) for the next step's
guidance_eps = guidance_eps + u_sw # sliding surface. The paper stores corrected guidance, but in
# ComfyUI the corrections accumulate through the surface's
# Store corrected guidance for next step's sliding surface # lambda * prev_eps term (amplified 4x per step at lambda=5),
# overwhelming the actual guidance signal after a few steps.
# Storing raw guidance keeps the surface tracking the model's
# actual guidance evolution while applying corrections fresh.
state["prev_eps"] = guidance_eps.detach().clone() state["prev_eps"] = guidance_eps.detach().clone()
# Convert back to sigma-scaled space and apply CFG # Apply correction and convert back to sigma-scaled space
return uncond + cond_scale * guidance_eps * sigma_val return uncond + cond_scale * (guidance_eps + u_sw) * sigma_val
m = model.clone() m = model.clone()
m.set_model_sampler_cfg_function(smc_cfg_function, disable_cfg1_optimization=True) m.set_model_sampler_cfg_function(smc_cfg_function, disable_cfg1_optimization=True)