Abstract
As one of the most popular and sought-after generative models in the recentyears, diffusion models have sparked the interests of many researchers andsteadily shown excellent advantage in various generative tasks such as imagesynthesis, video generation, molecule design, 3D scene rendering and multimodalgeneration, relying on their dense theoretical principles and reliableapplication practices. The remarkable success of these recent efforts ondiffusion models comes largely from progressive design principles and efficientarchitecture, training, inference, and deployment methodologies. However, therehas not been a comprehensive and in-depth review to summarize these principlesand practices to help the rapid understanding and application of diffusionmodels. In this survey, we provide a new efficiency-oriented perspective onthese existing efforts, which mainly focuses on the profound principles andefficient practices in architecture designs, model training, fast inference andreliable deployment, to guide further theoretical research, algorithm migrationand model application for new scenarios in a reader-friendly way.\url{https://github.com/ponyzym/Efficient-DMs-Survey}