[Paper Review] Diffusion Models Beat GANs on Image Synthesis