Use sqrt(cfg) scaling for K + store raw prev_eps

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>
This commit is contained in:
2026-03-04 23:39:16 +01:00
parent 2c50da025d
commit d29808975f

View File

@@ -37,24 +37,17 @@ class SMCCFGCtrl:
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,
}
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"] # x - cond_denoised (sigma-scaled noise)
uncond = args["uncond"] # x - uncond_denoised (sigma-scaled noise)
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.
# Detect new generation (sigma jumps up = new sampling run)
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
@@ -64,21 +57,13 @@ class SMCCFGCtrl:
step = state["step"]
state["step"] = step + 1
# Warmup: pure conditional prediction (no guidance)
if warmup_steps > 0 and step < warmup_steps:
return cond
# Normalize to noise-prediction space by dividing out sigma.
# The paper's K is calibrated for unit-variance noise predictions.
# ComfyUI's cond/uncond are (x - denoised) ≈ sigma * epsilon,
# so dividing by sigma recovers epsilon-space where K=0.2 is correct.
# Crucially, when converting back, the sigma factor naturally dampens
# the correction at late steps (small sigma), preventing noise injection.
# Normalize to noise-prediction space (divide out sigma).
sigma_val = max(curr_sigma, 1e-8)
guidance_eps = (cond - uncond) / sigma_val
# Initialize prev_eps on first SMC step (matches original paper
# where SMC correction is applied from the very first step)
if state["prev_eps"] is None:
state["prev_eps"] = guidance_eps.detach().clone()
@@ -87,29 +72,26 @@ class SMCCFGCtrl:
# Sliding surface: s_t = (e_t - e_{t-1}) + lambda * e_{t-1}
s = (guidance_eps - prev_eps) + lam * prev_eps
# Smooth switching via tanh(s/phi) instead of hard sign(s).
# The paper uses sign(s) which works in DiffSynth but creates
# salt-and-pepper artifacts in ComfyUI's latent space. tanh
# provides smooth spatial gradients: proportional near zero,
# saturating at ±K for large |s|.
phi = s.std().clamp(min=1e-6)
u_sw = -K * torch.tanh(s / phi)
# Scale K so the output correction (cond_scale * K_eff) stays
# in the range the paper tested (0.51.5). The paper only
# tested up to cfg=7.5; at cfg=12, full K gives cond_scale*K=2.4
# which corrupts the image. sqrt scaling keeps the output
# correction growing sub-linearly with cfg.
K_eff = K / max(cond_scale, 1.0) ** 0.5
# Spatial smoothing: blur the correction to remove per-element
# grid artifacts at VAE patch boundaries (each latent = 8x8 px).
# Smooth switching via tanh instead of sign to avoid
# salt-and-pepper artifacts in ComfyUI's latent space.
phi = s.std().clamp(min=1e-6)
u_sw = -K_eff * torch.tanh(s / phi)
# Spatial blur to smooth grid artifacts at VAE patch boundaries.
if u_sw.ndim == 4:
u_sw = F.avg_pool2d(u_sw, kernel_size=5, stride=1, padding=2)
# Store RAW guidance (before correction) for the next step's
# sliding surface. The paper stores corrected guidance, but in
# ComfyUI the corrections accumulate through the surface's
# 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.
# Store RAW guidance (before correction) to prevent correction
# accumulation through the lambda * prev_eps term.
state["prev_eps"] = guidance_eps.detach().clone()
# Apply correction and convert back to sigma-scaled space
return uncond + cond_scale * (guidance_eps + u_sw) * sigma_val
m = model.clone()