Y our Diffusion Model is Secretly a Noise Classifier and Benefits from Contrastive Training

Neural Information Processing Systems 

Diffusion models learn to denoise data and the trained denoiser is then used to generate new samples from the data distribution.