Nifty Wan Video SLG

WAN-native Spatiotemporal Skip Guidance (STG/SLG) — no TeaCache required. STG-A (default): zeros self-attention contribution only — less perturbation, lower flickering risk. STG-R: entire block becomes identity — stronger. Works at CFG=1 (LightX2V). Paper: arxiv 2411.18664 (CVPR 2025).

By Stibo·Created 5 months ago·Updated 2 months ago· 7

Nifty Wan Video SLG

model

MODEL

◄enabledtrue►

◄blocks9, 10►

◄mode▾►

◄scale1.00►

◄start_percent0.000►

◄end_percent1.000►

◄rescaling_scale0.00►

◄scale_end0.00►

Categorynifty/model

Inputs (9)

Name	Type	Default	Description
enabled	BOOLEAN	true	When false, node is a no-op and passes the input model unchanged.
model	MODEL		—
blocks	STRING	9, 10	Comma-separated block indices. Supports ranges: '9-11' = 9, 10, 11. WAN 2.2 = 40 layers. Blocks 9, 10 empirically recommended. Fewer blocks = safer. Start with '10' alone if flickering occurs.
mode	COMBO		STG-A (default): zeros self-attention only — cross-attn and FFN still run. Less perturbation, lower flickering risk. Recommended for WAN + few-step. STG-R: entire block becomes identity. Stronger but more flickering at 2-3 steps.
scale	FLOAT	1.000–10	Guidance strength. Works at CFG=1. LightX2V (2-3 HN steps): STG-A max ~2.0–2.5, STG-R max ~1.5. Use rescaling_scale to push higher without flickering.
start_percent	FLOAT	0.0000–1	Start of active range as fraction of the FULL denoising schedule (HN + LN samplers combined). Percentages are ABSOLUTE.
end_percent	FLOAT	1.0000–1	End of active range. Formula: HN_steps / total_steps. 4-step (2+2): 0.5. 6-step (3+3): 0.5. Default 1.0 = active for all sigmas this model processes.
rescaling_scaleopt	FLOAT	0.000–1	Normalizes correction std to match cond_pred std before scale. 0.0 = disabled. 0.3–0.5 = recommended if scale causes flickering. 1.0 = fully normalized.
scale_endopt	FLOAT	0.000–10	Sigma-adaptive scale: linearly interpolates from 'scale' (at sigma_start) to 'scale_end' (at sigma_end). 0.0 = disabled (constant scale). Example: scale=2.0, scale_end=0.5 — stronger guidance at high noise, fading toward low noise. Useful since STG has most structural impact early.

Outputs (1)

Name	Type	Description
MODEL	MODEL	—