ComfyUI-ScheduledGuider-Ext

Sheduled Guider Extension for ComfyUI

This extension contains various nodes for CFG scheduling.

SheduledCFGGuider Node

Illustration of SheduledCFGGuider Node interface

The SheduledCFGGuider node is designed to guide the sampling process using a scheduled CFG (Classifier-Free Guidance) strategy. It adjusts the guidance scale dynamically based on the noise schedule during diffusion.

Thanks to Clybius. I used his WarmupDecayCFGGuider's source code as a base.

Parameters

• model: The generative model to be used.
• positive: A conditioning input that specifies the desired attributes or style.
• unconditional: An unconditional input that serves as the base reference.
• cfg_max: The maximum value for the Classifier-Free Guidance scale (default: 12.0).
• cfg_min: The minimum value for the Classifier-Free Guidance scale (default: 1.0).
• sigmas: A list of sigma values defining the CFG-schedule.

Functionality

• Dynamically calculates the CFG value based on the current timestep and CFG-schedule.

Details

Current CFG value is mapped by sigmas input and scaled between cfg_min and cfg_max, based on the position of the current timestep (denoising sigma).

Increasing steps in sigmas input increase precision on CFG-scheduler curve, while keeping sampling steps count unchanged. You can set like 200 steps in your CFG scheduler. Guider will compute current CFG value depending on your denoising schedule.

PerpNegSheduledCFGGuider Node

Illustration of PerpNegSheduledCFGGuider Node interface

The PerpNegSheduledCFGGuider node extends the functionality of SheduledCFGGuider by incorporating negative conditioning. This allows for more nuanced control over the generated output by penalizing unwanted features.

Parameters

• model: The generative model to be used.
• positive: A conditioning input specifying desired attributes or style.
• negative: A conditioning input specifying unwanted features.
• unconditional: An unconditional input serving as the base reference.
• cfg_max: Maximum value for the Classifier-Free Guidance scale (default: 12.0).
• cfg_min: Minimum value for the Classifier-Free Guidance scale (default: 1.0).
• neg_scale: Scaling factor for the negative conditioning (default: 1.0).
• sigmas: List of sigma values defining the CFG-schedule.
• use_negative_as_unconditional: Boolean flag to determine if the negative condition should act as an unconditional baseline on CFG post processing (default: True).

CosineScheduler Node

Illustration of CosineScheduler Node interface

The CosineScheduler node generates a schedule of sigma values using a cosine function. This scheduler provides a smooth transition between defined maximum and minimum sigma values over a specified number of steps for specified period of cosine function.

Parameters

• steps (INT): Number of discrete steps in the schedule. Determines how many sigma values will be generated.

  • Default: 20
  • Range: 1 to 10000

• sigma_max (FLOAT): The maximum value of sigma in the schedule.

  • Default: 1.0
  • Range: 0.0 to 5000.0

• sigma_min (FLOAT): The minimum value of sigma in the schedule.

  • Default: 0.0
  • Range: 0.0 to 5000.0

• period (FLOAT): Controls the periodicity of the cosine function applied to generate sigma values.

  • Default: 0.5
  • Range: 0.0 to 5000.0

Functionality

• Computes a list of sigma values using a cosine function that transitions smoothly from sigma_max to sigma_min.
• Utilizes a cosine curve modulated by the specified period to adjust the rate and nature of this transition.

GaussianScheduler Node

The GaussianScheduler node generates a schedule