Masked Diffusion Models are Fast Learners

Diffusion models have emerged as the de-facto technique for image generation, yet they entail significant computational overhead, hindering the technique's broader application in the research community. We propose a prior-based denoising training framework, the first to incorporate the pre-train and fine-tune paradigm into the diffusion model training process, which substantially improves training efficiency and shows potential in facilitating various downstream tasks. Our approach centers on masking a high proportion (e.g., up to 90%) of the input image and employing masked score matching to denoise the visible areas, thereby guiding the diffusion model to learn more salient features from training data as prior knowledge. By utilizing this masked learning process in a pre-training stage, we efficiently train the ViT-based diffusion model on CelebA-HQ 256x256 in the pixel space, achieving a 4x acceleration and enhancing the quality of generated images compared to DDPM. Moreover, our masked pre-training technique is universally applicable to various diffusion models that directly generate images in the pixel space and facilitates learning pre-trained models with excellent generalizability: a diffusion model pre-trained on VGGFace2 attains a 46% quality improvement through fine-tuning with merely 10% local data. Our code is available at https://github.com/jiachenlei/maskdm.