ComfyUI Node: Wan MoE KSampler

Authored by stduhpf

Created

Updated

126 stars

Category

sampling

Inputs

model_high_noise MODEL
model_low_noise MODEL
boundary FLOAT
seed INT
steps INT
cfg_high_noise FLOAT
cfg_low_noise FLOAT
sampler_name
  • euler
  • euler_cfg_pp
  • euler_ancestral
  • euler_ancestral_cfg_pp
  • heun
  • heunpp2
  • dpm_2
  • dpm_2_ancestral
  • lms
  • dpm_fast
  • dpm_adaptive
  • dpmpp_2s_ancestral
  • dpmpp_2s_ancestral_cfg_pp
  • dpmpp_sde
  • dpmpp_sde_gpu
  • dpmpp_2m
  • dpmpp_2m_cfg_pp
  • dpmpp_2m_sde
  • dpmpp_2m_sde_gpu
  • dpmpp_2m_sde_heun
  • dpmpp_2m_sde_heun_gpu
  • dpmpp_3m_sde
  • dpmpp_3m_sde_gpu
  • ddpm
  • lcm
  • ipndm
  • ipndm_v
  • deis
  • res_multistep
  • res_multistep_cfg_pp
  • res_multistep_ancestral
  • res_multistep_ancestral_cfg_pp
  • gradient_estimation
  • gradient_estimation_cfg_pp
  • er_sde
  • seeds_2
  • seeds_3
  • sa_solver
  • sa_solver_pece
  • ddim
  • uni_pc
  • uni_pc_bh2
scheduler
  • simple
  • sgm_uniform
  • karras
  • exponential
  • ddim_uniform
  • beta
  • normal
  • linear_quadratic
  • kl_optimal
sigma_shift FLOAT
positive CONDITIONING
negative CONDITIONING
latent_image LATENT
denoise FLOAT

Outputs

LATENT

Extension: KSampler for Wan 2.2 MoE for ComfyUI

These nodes are made to support 'Mixture of Expert' Flow models with the architecture of Wan2.2 A14B (With a high noise expert and low noise expert). Instead of guessing the denoising step at which to swap from tyhe high noise model to the low noise model, this node automatically chanage to the low noise model when we reach the diffusion timestep at which the signal to noise ratio is supposed to be 1:1.

Authored by stduhpf

Run ComfyUI workflows in the Cloud!

No downloads or installs are required. Pay only for active GPU usage, not idle time. No complex setups and dependency issues

Learn more