Masked Diffusion Models Are Fast Distribution Learners

Diffusion models have emerged as the \emph{de-facto} generative model for image synthesis, yet they entail significant training overhead, hindering the technique's broader adoption in the research community. We observe that these models are commonly trained to learn all fine-grained visual information from scratch, thus motivating our investigation on its necessity. In this work, we show that it suffices to set up pre-training stage to initialize a diffusion model by encouraging it to learn some primer distribution of the unknown real image distribution. Then the pre-trained model can be fine-tuned for specific generation tasks efficiently. To approximate the primer distribution, our approach centers on masking a high proportion (e.g., up to 90\%) of an input image and employing masked denoising score matching to denoise visible areas. Utilizing the learned primer distribution in subsequent fine-tuning, we efficiently train a ViT-based diffusion model on CelebA-HQ $256 \times 256$ in the raw pixel space, achieving superior training acceleration compared to denoising diffusion probabilistic model (DDPM) counterpart and a new FID score record of 6.73 for ViT-based diffusion models. Moreover, our masked pre-training technique can be universally applied to various diffusion models that directly generate images in the pixel space, aiding in the learning of pre-trained models with superior generalizability. For instance, a diffusion model pre-trained on VGGFace2 attains a 46\% quality improvement through fine-tuning on only 10\% data from a different dataset. Our code is available at \url{https://github.com/jiachenlei/maskdm}.