Nodes/Hermes Nodes/WAN Dual KSampler
ComfyUI Node

WAN Dual KSampler

Dual KSampler for WAN 2.2 MoE: runs HIGH model for initial steps, then LOW model for remaining steps.

By Lannister34·Created 10 months ago·Updated 10 days ago· 1
WAN Dual KSampler
  • model_high
  • model_low
  • positive
  • negative
  • latent_image
  • latent
◄seed0►
◄steps8►
◄switch_step4►
◄first_cfg6.0►
◄second_cfg6.0►
◄sampler_name▾►
◄scheduler▾►
◄denoise1.00►
◄offload_between_phasesfalse►
Categorysampling

Inputs (14)

NameTypeDefaultDescription
model_highMODELThe HIGH noise model (handles initial denoising steps).
model_lowMODELThe LOW noise model (handles final denoising steps).
seedINT00–18446744073709550000The random seed used for creating the noise.
stepsINT81–10000Total number of denoising steps.
switch_stepINT41–10000Step at which to switch from HIGH to LOW model.
first_cfgFLOAT6.00–100CFG scale for HIGH model phase.
second_cfgFLOAT6.00–100CFG scale for LOW model phase.
sampler_nameCOMBOThe sampling algorithm to use.
schedulerCOMBOThe scheduler for noise removal.
positiveCONDITIONINGPositive conditioning (prompt).
negativeCONDITIONINGNegative conditioning.
latent_imageLATENTThe latent image to denoise.
denoiseFLOAT1.000–1Denoising strength.
offload_between_phasesBOOLEANfalseUnload models from VRAM between the HIGH and LOW phases and again after sampling. Helps on low-VRAM GPUs at the cost of reload time.

Outputs (1)

NameTypeDescription
latentLATENTThe denoised latent.