skrample 0.1.0

Composable sampling functions for diffusion models

skrample

Composable sampling functions for diffusion models

Status

Vertical slice, gradually overtaking many diffusers features in quickdif

Feature Flags

• beta-schedule -> scipy : For the Beta() schedule modifier
• brownian-noise -> torchsde : For the Brownian() noise generator
• 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
  • Ancestral
• DPM
  • 1st order, 2nd order, 3rd order
  • SDE
• IPNDM
  • Ancestral (from Euler)
• UniPC
  • N order, limited to 9 for stability
  • Custom solver via other SkrampleSampler types

Schedules

• Linear
• Scaled
  • uniform flag, AKA "trailing" in diffusers
• Flow
  • Dynamic and non-dynamic shifting
• ZSNR

Schedule modifiers

• Karras
• Exponential
• Beta

Predictors

• Epsilon
• Velocity / vpred
• Flow

Noise generators

• Random
• Brownian
• Offset
• Pyramid

Integrations

Diffusers

• Compatibility for pipelines
  • SD1
  • SDXL
  • SD3
  • Flux
  • Others?
• Import from config
  • Sampler
    • Not sure this is even worthwhile. All Skrample samplers work everywhere
  • Schedule
  • Predictor
• Manage state
  • Steps
  • Higher order
  • Generators
  • Config as presented

Implementations

quickdif

A basic test bed is available for https://github.com/Beinsezii/quickdif.git on the skrample branch