Nodes/Nifty Nodes for ComfyUI/Wan Video NAG
ComfyUI Node

Wan Video NAG

Normalized Attention Guidance (NAG) for WAN video models (T2V and I2V, 2.1 and 2.2). Replaces or complements CFG by applying guidance directly in attention space — stable across few-step and multi-step sampling. Recommended starting values — T2V multi-step: scale=11, tau=2.5, alpha=0.25 | T2V few-step (Lightning): scale=11-15, tau=2.5-5, alpha=0.25-0.5 | I2V (any): scale=11, tau=1.5-2.0, alpha=0.10-0.20 (lower to preserve reference image)

By Stibo·Created 5 months ago·Updated 2 months ago· 7
Wan Video NAG
  • model
  • conditioning
  • MODEL
  • CONDITIONING
◄enabledtrue►
◄nag_scale11.000►
◄nag_alpha0.250►
◄nag_tau2.500►
◄nag_sigma_end0.00►
◄input_type▾►
◄inplacefalse►
◄cond_zero_outtrue►
◄nag_scale_end0.000►
Categorynifty/model

Inputs (11)

NameTypeDefaultDescription
enabledBOOLEANtrueMaster switch. When false the node is a no-op: model and conditioning pass through unchanged.
modelMODEL—
conditioningCONDITIONINGYour NEGATIVE conditioning (empty or negative prompt). NAG uses this as its internal reference — the direction the model is guided AWAY from in attention space. Do NOT connect the positive conditioning here. With cond_zero_out=True, this also flows to the CONDITIONING output as a zeroed tensor for the sampler's negative slot.
nag_scaleFLOAT11.0000–100Guidance strength (φ), analogous to CFG scale. Good range for WAN: 8–15. Setting to 0 disables NAG. Tune this last — first settle on nag_tau and nag_alpha.
nag_alphaFLOAT0.2500–1Blend weight (α) between NAG-guided and original positive attention (Eq. 10). 0 = no effect, 1 = full NAG. T2V default: 0.25 | I2V: use lower (0.10–0.20) to preserve the reference image. Few-step (Lightning LoRA): can go higher (0.3–0.5).
nag_tauFLOAT2.5000–10L1-norm clipping threshold (τ): caps how far guided attention deviates from positive attention (Eq. 9 in paper). Lower = safer, less drift. Higher = stronger correction. T2V default: 2.5 | I2V: use lower (1.5–2.0) to preserve the reference image. Few-step (Lightning LoRA): can go higher (3–5). Paper ablation: both tau and alpha are critical — without them output degrades sharply above scale=5.
nag_sigma_endFLOAT0.000–1NAG is skipped when denoising sigma drops below this value. 0.0 = always active (correct for two-model high/low-noise pipelines). For single-model pipelines: 0.75 achieves near-identical quality with significantly less compute (paper authors' recommendation for flow-based models). Note: WAN uses Rectified Flow where sigma runs 1.0 → 0.0.
input_typeCOMBOdefault: sampler sends a [positive, negative] batch pair (standard CFG setup). batch: single conditioning without a paired negative — use when sampling without CFG (e.g. distilled/few-step WAN models with Lightning LoRA).
inplaceBOOLEANfalseModify tensors in-place to reduce peak VRAM. Slightly alters numerical results. Enable only if out of memory.
cond_zero_outBOOLEANtrueOutput a zeroed-out (neutral) conditioning instead of the input conditioning. Recommended: NAG handles guidance internally, so the sampler's negative slot should be empty to avoid interference. Disable only if intentionally stacking NAG on top of CFG.
nag_scale_endoptFLOAT0.0000–100Sigma-adaptive scale: linearly interpolates from 'nag_scale' (at sigma=1.0) to 'nag_scale_end' (at sigma=0.0). 0.0 = disabled (uses constant nag_scale).

Outputs (2)

NameTypeDescription
MODELMODEL—
CONDITIONINGCONDITIONING—