WaveDiff is a wavelet-based diffusion scheme for image generation that decomposes images into low-and-high frequency components using wavelet transforms at both image and feature levels. The approach adaptively accelerates the sampling process while maintaining generation quality. The implementation includes training and evaluation code for datasets including CelebA-HQ, CIFAR-10, LSUN-Church, and STL-10, targeting state-of-the-art inference speed for diffusion models.