This paper presents improved techniques for training score-based generative models, which can generate high-quality image samples comparable to GANs without requiring adversarial optimization. The key challenges in training these models include instability and limited resolution, as existing methods typically work only on low-resolution images (e.g., 32x32). The authors propose a new theoretical analysis of learning and sampling from score-based models in high-dimensional spaces, which helps explain existing failure modes and motivates new solutions that generalize across datasets. They also introduce an exponential moving average of model weights to enhance stability. The proposed techniques allow score-based generative models to scale to various image datasets with resolutions ranging from 64x64 to 256x256. These models can generate high-fidelity samples that rival the best-in-class GANs on datasets such as CelebA, FFHQ, and several LSUN categories. The techniques include choosing appropriate noise scales, incorporating noise information, and configuring annealed Langevin dynamics. The authors also introduce an exponential moving average (EMA) to improve stability during training. The paper demonstrates that the proposed techniques significantly improve the performance of score-based generative models, leading to better sample quality and more stable training. The results show that the models can generate high-fidelity images with resolutions up to 256x256, which was previously impossible with score-based generative models. The techniques are effective across a wide range of image datasets and have been validated through extensive experiments. The authors conclude that their techniques significantly improve the training and sampling processes of score-based generative models, leading to better sample quality and enabling high-fidelity image generation at high resolutions.This paper presents improved techniques for training score-based generative models, which can generate high-quality image samples comparable to GANs without requiring adversarial optimization. The key challenges in training these models include instability and limited resolution, as existing methods typically work only on low-resolution images (e.g., 32x32). The authors propose a new theoretical analysis of learning and sampling from score-based models in high-dimensional spaces, which helps explain existing failure modes and motivates new solutions that generalize across datasets. They also introduce an exponential moving average of model weights to enhance stability. The proposed techniques allow score-based generative models to scale to various image datasets with resolutions ranging from 64x64 to 256x256. These models can generate high-fidelity samples that rival the best-in-class GANs on datasets such as CelebA, FFHQ, and several LSUN categories. The techniques include choosing appropriate noise scales, incorporating noise information, and configuring annealed Langevin dynamics. The authors also introduce an exponential moving average (EMA) to improve stability during training. The paper demonstrates that the proposed techniques significantly improve the performance of score-based generative models, leading to better sample quality and more stable training. The results show that the models can generate high-fidelity images with resolutions up to 256x256, which was previously impossible with score-based generative models. The techniques are effective across a wide range of image datasets and have been validated through extensive experiments. The authors conclude that their techniques significantly improve the training and sampling processes of score-based generative models, leading to better sample quality and enabling high-fidelity image generation at high resolutions.