This repository implements the score-based generative modeling framework through stochastic differential equations (SDEs), originally published at ICLR 2021. The method transforms data into noise via a continuous-time stochastic process, then reverses the SDE for sample generation using score matching. It achieves state-of-the-art FID of 2.20 on CIFAR-10 and generates high-fidelity 1024px Celeba-HQ images. The work supports various applications including class-conditional generation, inpainting, and colorization.