Composable sampling functions for diffusion models

Production-tested on all popular diffusion models. The library has significantly matured since 0.5

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()

• beta-schedule -> scipy : For the Beta() schedule modifier
• brownian-noise -> torchsde : For the Brownian() noise generator
• cdf-schedule -> scipy : For the Probit() 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/

These samplers are written inside-out to be compatible with Diffusers and similar frameworks

• Euler
  • Stochastic
• DPM
  • Order 1-3
  • Stochastic
• Adams/IPNDM
  • Order 1-9
  • Stochastic
• UniP & UniPC
  • Order 1-9
  • Stochastic
  • Custom predictor via other SkrampleSampler types
• SPC
  • Basic fully customizable midpoint corrector

These samplers are written using closures similar to ksampler

• RKUltra
  • Arbitrary Runge-Kutta solver
  • Order 1-15, customizable through tableaux system
  • Stochastic
• DynasauRK
  • Procedural Runge-Kutta solver
  • Order 2-4
  • Stochastic
• RKMoire
  • Experimental
  • Embedded Runge-Kutta solver
  • Order 2-6, customizable through tableaux system

• Linear
  • Flow-matching default
• Scaled
  • Variance-preserving default
• ZSNR

Replaces sigmas on an existing schedule

• Karras
• Exponential
• Beta
• Probit

Modifies timestep spacing of a schedule

• FlowShift
• Hyper
• Sinner

• Data / Sample / X-Pred
• Noise / Epsilon / Ε-Pred
• Velocity / V-Pred
• Flow / U-pred

• Random
• Brownian
• Offset
• Pyramid

• Compatibile with DiffusionPipeline
• Import from config
  • Sampler
  • Schedule
  • Predictor
• Structured sampler wrapper
• Functional sampler wrappers
  • RKUltra
  • DynasauRK

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