Skrample 0.4.2

Composable sampling functions for diffusion models

Status

Mostly complete for common models, superseding all diffusers features in quickdif

Quickstart

Fastest way to jump in is examples. The classes and functions themselves have docstrings and type hints, so it's recommended to make liberal use of your IDE or python help()

Feature Flags

• beta-schedule -> scipy : For the Beta() schedule modifier
• brownian-noise -> torchsde : For the Brownian() noise generator
• cdf-schedule -> scipy : For the SigmoidCDF() schedule
• diffusers-wrapper -> torch : For the diffusers integration module
• pytorch -> torch : For the pytorch module
  • pytorch.noise : Custom generators
• all : All of the above
• dev : For running tests/

Samplers

• Euler
• DPM
  • 1st order, 2nd order, 3rd order
  • SDE
• Adams/IPNDM
• UniP & UniPC
  • N order, limited to 9 for stability
  • Custom solver via other SkrampleSampler types
• SPC
  • Basic fully customizable midpoint corrector

Schedules

• Linear
• Scaled
  • uniform flag, AKA "trailing" in diffusers
• SigmaCDF
• ZSNR

Schedule modifiers

• Karras
• Exponential
• FlowShift
• Beta
• Hyper

Predictors

• Epsilon
• Velocity / vpred
• Flow

Noise generators

• Random
• Brownian
• Offset
• Pyramid

Integrations

Diffusers

• Compatibility for pipelines
  • SD1
  • SDXL
  • SD3
  • Flux
• Import from config
  • Sampler
  • Schedule
  • Predictor
• Manage state
  • Steps
  • Higher order
  • Generators
  • Config as presented

Implementations

quickdif

My diffusers cli quickdif has full support for all major Skrample features, allowing extremely fine-grained customization.