1: \begin{abstract}

2: A potent class of generative models known as Diffusion Probabilistic Models

3: (DPMs) has become prominent. A forward diffusion process adds gradually noise

4: to data, while a model learns to gradually denoise. Sampling from pre-trained

5: DPMs is obtained by solving differential equations (DE) defined by the learnt

6: model, a process which has shown to be prohibitively slow. Numerous efforts on

7: speeding-up this process have consisted on crafting powerful ODE solvers.

8: Despite being quick, such solvers do not usually reach the optimal quality

9: achieved by available slow SDE solvers. Our goal is to propose SDE solvers that

10: reach optimal quality without requiring several hundreds or thousands of NFEs

11: to achieve that goal. In this work, we propose Stochastic Exponential

12: Derivative-free Solvers (SEEDS), improving and generalizing Exponential

13: Integrator approaches to the stochastic case on several frameworks. 

14: After carefully analyzing the formulation of exact

15: solutions of diffusion SDEs, we craft SEEDS to analytically compute the linear

16: part of such solutions. Inspired by the Exponential Time-Differencing method,

17: SEEDS uses a novel treatment of the stochastic components of solutions,

18: enabling the analytical computation of their variance, and contains high-order

19: terms allowing to reach optimal quality sampling $\sim3$-$5\times$ faster than previous

20: SDE methods. We validate our approach on several image generation benchmarks,

21: showing that SEEDS outperforms or is competitive with previous SDE solvers.

22: Contrary to the latter, SEEDS are derivative and training free, and we fully

23: prove strong convergence guarantees for them.

24:  \end{abstract}

25: