Score-based Generative Models (SGMs) have been considered by some as the new competitor to GANs in Generative Modeling. They have been able to give high quality samples without any adversarial sampling, while also providing the facility to compute exact likelihood. In this project, we propose a further improvement in their score estimation framework by utilizing multiple noisy inputs instead of just one. We build our idea over the well-known equivalence between score matching and denoising autoencoders and show that the model along with Tweedie’s formula can be easily extended to multiple samples, thus giving it the capacity to obtain better score estimates close to true distribution scores for langevin dynamics sampling.
