DPMSolverMultistepScheduler<T>

Models & Types · AiDotNet.Diffusion.Schedulers

DPM-Solver++ multistep scheduler for fast diffusion model sampling.

Inheritance: NoiseSchedulerBase<T> → DPMSolverMultistepScheduler<T>

For Beginners

DPM-Solver++ is one of the fastest high-quality samplers.

Think of it like navigating with a GPS that remembers your path:

  • Euler: Looks at current position only, takes simple steps
  • DPM-Solver++: Remembers previous positions, predicts better next steps

Key characteristics:

  • Very fast: Good quality in just 15-25 steps
  • Multi-order: Uses 1st, 2nd, or 3rd order methods adaptively
  • Deterministic: Same seed always gives same result
  • Widely adopted: Used as default in many Stable Diffusion implementations

The “multistep” means it stores previous model outputs and uses them to make more accurate predictions, similar to Adams-Bashforth methods in numerical analysis.

How It Works

DPM-Solver++ is a high-order ODE solver specifically designed for diffusion models. It achieves state-of-the-art sampling quality with very few steps (10-25) by using multi-step methods that leverage history of previous model evaluations.

Reference: Lu et al., “DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models”, 2022

Constructors

Constructor Summary
DPMSolverMultistepScheduler(SchedulerConfig<T> config, int solverOrder) Initializes a new instance of the DPM-Solver++ multistep scheduler.

Methods

Method Summary
Dictionary<string, object> GetState()
Vector<T> Step(Vector<T> modelOutput, int timestep, Vector<T> sample, T eta, Vector<T> noise) Performs one DPM-Solver++ multistep denoising step.
void LoadState(Dictionary<string, object> state)
void SetTimesteps(int inferenceSteps) Sets up the inference timesteps and computes lambda/sigma/alpha schedules.