1: \begin{abstract}

2: Score-based generative models (SGMs) have recently emerged as a promising class of generative models.

3: The key idea is to produce high-quality images by recurrently adding Gaussian noises and gradients to a Gaussian sample until converging to the target distribution, a.k.a. the diffusion sampling.

4: %

5: To ensure stability of convergence in sampling and generation quality, however, this sequential sampling process has to take a {\em small step size} and many sampling iterations (\eg, 2000).

6: %

7: Several acceleration methods have been proposed

8: with focus on low-resolution generation. 

9: In this work, we consider the acceleration of high-resolution 

10: generation with SGMs, a more challenging yet more important problem. 

11: We prove theoretically that this {slow convergence} drawback is primarily due to the ignorance of the target distribution.

12: Further, we introduce a novel {\bf\em Target Distribution Aware Sampling} (TDAS) method by leveraging the structural priors in space and frequency domains.

13: Extensive experiments on CIFAR-10, CelebA, LSUN, and FFHQ datasets validate that TDAS can consistently accelerate state-of-the-art SGMs, particularly on more challenging high resolution ($1024\times 1024$) image generation tasks by up to $18.4\times$, whilst largely maintaining the synthesis quality.

14: With fewer sampling iterations, TDAS can still generate good quality images. In contrast, the existing methods degrade drastically or even fails completely.

15: \end{abstract}

16: