ComfyUI-RK-Sampler

Installation

ComfyUI-Manager

ComfyUI Manager Menu > Custom Nodes Manager > ComfyUI-RK-Sampler > Install

Manual installation

From ComfyUI/custom_nodes and ComfyUI virtual environment:

git clone https://github.com/wootwootwootwoot/ComfyUI-RK-Sampler.git
pip install torchode

Usage

workflow Basic workflow

From Add Node:

sampling > custom_sampling > samplers > Runge-Kutta Sampler

Best defaults

• These methods support normal as well as high CFG scales, but the results may depend on the specific model.
• For SDXL, the best results I got were from CFG scales between 7-35.
• If you don't know the right step count or CFG scale:
  • Try the adaptive_pid controller with the base CFG and increment it until the results get worse.
  • Tune the CFG to the desired output.
  • Try the fixed_scheduled controller:
    • Use the Align Your Steps scheduler.
    • Set the scheduler's step count to be the same as the number of steps taken by the adaptive_pid controller.
    • Leave the CFG scale unchanged.
  • Tune the scheduler step count.
• When using adaptive_pid or adaptive_scipy controllers:
  • Set log_absolute_tolerance to -3.5:
    • Decrease it to set more strict tolerances (better results) in exchange for slower inference times.
    • Increase it to set less strict tolerances (worse results) in exchange for faster inference times.
  • Set log_relative_tolerance to be 1-2 more than log_absolute_tolerance.

Adaptive step size

method: ae_bosh3
step_size_controller: adaptive_pid
log_absolute_tolerance: -3.5
log_relative_tolerance: -2.5
pcoeff: 0
icoeff: 1
dcoeff: 0
norm: rms_norm
enable_dt_min: false
enable_dt_max: true
dt_min: -0.1
dt_max: 0
safety: 0.9
factor_min: 0.2
factor_max: 10
max_steps: 2147483647
min_sigma: 1e-5
cfg: 7-35

Fixed step size

method: fe_ralston3
step_size_controller: fixed_scheduled
scheduler: Align Your Steps
steps: 28-150
cfg: 7-35

Choose a step size controller

step_size_controller: Controller to determine the step size taken on each sampling step.

adaptive controllers: Automatically determines the step size(s). The scheduler choice and scheduler step count does not matter since they only use the start and end timesteps.

fixed controllers: Uses the step size(s) provided by the scheduler. Works like a normal non-adaptive sampler in this case.

• adaptive_pid: A proportional–integral–derivative (PID) controller. Works with a-class methods.
• fixed_scheduled: A controller that uses the $\sigma$ (timestep) schedule from the scheduler. Works with a-class and f-class methods.
• adaptive_scipy: A basic integral controller wrapped from scipy. Works with s-class methods.

Choose a method

method: Determines the solver method used.

a, f, and s classes

• a = adaptive, f = fixed, s = scipy, e = explicit
• Use a-class methods with either:
  • adaptive_pid for automatically determined step sizes/count.
  • fixed_scheduled for scheduler determined step sizes/count.
• Use f-class methods with fixed_scheduled.
• Use s-class methods with adaptive_scipy.

Recommended methods

• Try ae_bosh3, ae_dopri5, and ae_fehlberg5 with the adaptive_pid step size controller.
• Try fe_ralston3, ae_bosh3, and fe_ssprk3 with the fixed_scheduled step size controller.

Verificado 3 ago