This paper presents a comprehensive analysis of the design space of diffusion-based generative models, aiming to simplify and clarify the design choices involved in both training and sampling processes. The authors argue that current diffusion models are overly complex and propose a modular framework that separates concrete design decisions, enabling more effective improvements. Their approach leads to significant improvements in image quality, as evidenced by state-of-the-art FID scores of 1.79 for CIFAR-10 and 1.97 for unconditional settings, with faster sampling (35 network evaluations per image) compared to prior designs. They also demonstrate that their design changes significantly improve the FID of a previously trained ImageNet-64 model, reducing it from 2.07 to 1.55 and further to 1.36 after re-training with their improvements. The paper introduces a common framework for diffusion models, expressing them in terms of a probability flow ODE and analyzing the impact of different scheduling choices on sampling efficiency and quality. They propose a 2nd-order deterministic sampling method that significantly reduces the number of sampling steps required, leading to faster and more efficient sampling. They also explore stochastic sampling, showing that while it can introduce noise, it is not always beneficial and can sometimes be harmful, depending on the model and dataset. The authors also address the training of score-modeling neural networks, proposing a principled analysis of input, output, and loss function preconditioning, and suggest an improved distribution of noise levels during training. They also introduce a new approach to data augmentation that improves the quality of generated images, leading to new state-of-the-art results on CIFAR-10. Overall, the paper provides a detailed analysis of the design space of diffusion-based generative models, offering practical insights and improvements that enhance the performance and efficiency of these models. The authors believe that their approach will enable more extensive and targeted exploration of the design space of diffusion models, leading to further advancements in the field.This paper presents a comprehensive analysis of the design space of diffusion-based generative models, aiming to simplify and clarify the design choices involved in both training and sampling processes. The authors argue that current diffusion models are overly complex and propose a modular framework that separates concrete design decisions, enabling more effective improvements. Their approach leads to significant improvements in image quality, as evidenced by state-of-the-art FID scores of 1.79 for CIFAR-10 and 1.97 for unconditional settings, with faster sampling (35 network evaluations per image) compared to prior designs. They also demonstrate that their design changes significantly improve the FID of a previously trained ImageNet-64 model, reducing it from 2.07 to 1.55 and further to 1.36 after re-training with their improvements. The paper introduces a common framework for diffusion models, expressing them in terms of a probability flow ODE and analyzing the impact of different scheduling choices on sampling efficiency and quality. They propose a 2nd-order deterministic sampling method that significantly reduces the number of sampling steps required, leading to faster and more efficient sampling. They also explore stochastic sampling, showing that while it can introduce noise, it is not always beneficial and can sometimes be harmful, depending on the model and dataset. The authors also address the training of score-modeling neural networks, proposing a principled analysis of input, output, and loss function preconditioning, and suggest an improved distribution of noise levels during training. They also introduce a new approach to data augmentation that improves the quality of generated images, leading to new state-of-the-art results on CIFAR-10. Overall, the paper provides a detailed analysis of the design space of diffusion-based generative models, offering practical insights and improvements that enhance the performance and efficiency of these models. The authors believe that their approach will enable more extensive and targeted exploration of the design space of diffusion models, leading to further advancements in the field.