Skip to content

Configuration_Options.md

Latest commit

 

History

History
142 lines (85 loc) · 5.06 KB

Configuration_Options.md

File metadata and controls

142 lines (85 loc) · 5.06 KB

Simple Karras Exponential Scheduler (SimpleKEScheduler)

A flexible sigma scheduler for Stable Diffusion pipelines supporting:

  • Karras and Exponential scheduling
  • Advanced blending and noise scaling
  • Optional randomization (symmetric, asymmetric, or manual)
  • Auto sigma min/max scaling

Configuration Options

Key Description Type Default
sigma_min Minimum sigma value (can be randomized or auto-controlled) float Required
sigma_max Maximum sigma value (can be randomized or auto-controlled) float Required
rho Controls steepness of Karras curve float 7.0
sigma_auto_enabled Enable auto mode to control sigma_min or sigma_max bool False
sigma_auto_mode 'sigma_min' or 'sigma_max' to control which sigma is auto-scaled string 'sigma_min'
sigma_scale_factor Scale factor used when auto mode is enabled float 1000
global_randomize Enables randomization for all eligible parameters bool False
randomization_type 'asymmetric', 'symmetric', or 'off' (can override per-parameter) string 'asymmetric'
randomization_percent Global randomization range percentage float 0.2
debug Enables detailed logging bool False

Detailed Parameter Descriptions

Sigma Parameters

  • sigma_min and sigma_max

    • Define the minimum and maximum sigma values for noise scheduling.
    • Can be randomized or auto-controlled based on sigma_auto_enabled and sigma_auto_mode.
    • Auto mode recalculates one based on the other using sigma_scale_factor.

Auto Sigma Control

  • sigma_auto_enabled

    • If enabled, automatically controls either sigma_min or sigma_max.

  • sigma_auto_mode

    • Determines which sigma is auto-calculated: sigma_min or sigma_max.

  • sigma_scale_factor

    • Defines the scaling ratio when auto mode is active.

Randomization Controls

  • global_randomize

    • If enabled, activates randomization for all supported parameters (_rand_min and _rand_max values only)

  • randomization_type

    *_randomization_type

Specifies the type of advanced randomization.

Supported types:

symmetric → Random range is evenly distributed above and below the base value.

asymmetric → Random range is weighted with a larger upper range.

logarithmic → Random values are selected in log-space, useful for scaling-sensitive parameters.

exponential → Random values are generated using exponential growth based on a random base.

Short Name Aliases:

  • s, sym → symmetric

  • a, asym → asymmetric

  • l, log → logarithmic

  • e, exp → exponential

    randomization_percent

    • Defines the randomization range as a percentage of the base value.

Blending Controls

  • start_blend and end_blend

    • Control the dynamic blend factor between Karras and Exponential sigma schedules.
    • Blending is progress-based, transitioning smoothly from start_blend to end_blend over the steps.

Sharpness Control

  • sharpness

    • Applies an adaptive sharpening mask to the sigma sequence.
    • Increases detail retention when sigma values drop below a threshold.

Step Size Controls

  • initial_step_size and final_step_size

    • Define the step size at the beginning and end of the schedule.
    • Adjusted dynamically per step based on progress.

Noise Scaling Controls

  • initial_noise_scale and final_noise_scale

    • Define the noise scaling factor at the beginning and end of the schedule.
    • Adjusted dynamically per step based on progress.

Smooth Blending Control

  • smooth_blend_factor

    • Controls how smoothly the blend between Karras and Exponential schedules transitions.
    • Higher values create softer transitions.

Adaptive Step and Noise Controls

  • step_size_factor

    • Further scales the step size progression.

  • noise_scale_factor

    • Further scales the noise scaling progression.

Early Stopping Control

  • early_stopping_threshold

    • Defines the minimum change required between steps.
    • If all step changes fall below this threshold, the scheduler triggers early stopping to optimize computation.

Logging

  • The scheduler writes generation logs to modules/sd_simple_kes/image_generation_data.
  • Logs include full parameter tracking, sigma sequences, randomization details, and any auto mode adjustments.