Research implementation of DPM-Solver++ for fast sampling from diffusion probabilistic models.

• DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling (NeurIPS 2022)
• DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models (2022)

import torch
from dpm_solver import DPMSolver, model_wrapper

# Wrap your diffusion model
model_fn = model_wrapper(model, guidance_type="classifier-free", guidance_scale=7.5)

# Initialize solver
solver = DPMSolver(model_fn, algorithm_type="dpmsolver++")

# Sample
x_T = torch.randn(1, 3, 512, 512).cuda()
samples = solver.sample(x_T, steps=20, order=3, method='singlestep')

DPM-Solver++ converts the diffusion ODE:

dx/dt = (f(t)x + g(t)²∇_x log p_t(x)) / 2

Into log-SNR coordinates λ_t = log(α_t/σ_t) where the ODE becomes:

dx/dλ = (e^(-λ)σ_λ/α_λ)(α_λx_0_θ(x,λ) - x)

The solver uses exponential integrators with order 1, 2, or 3 for fast convergence.

pip install torch numpy

• Singlestep: High-order solver with intermediate evaluations
• Multistep: Adams-Bashforth style multistep method
• Adaptive: Automatic step size control with error estimation

• dpmsolver++: Data prediction variant (recommended)
• dpmsolver: Noise prediction variant