Jonathan Ho, Ajay Jain, & Pieter Abbeel (2020)
arXiv.
DOI: https://doi.org/10.48550/arxiv.2006.11239
Abstract. Introduces Denoising Diffusion Probabilistic Models (DDPM) in their modern form, showing that a simple noise-prediction training objective yields high-quality image generation. DDPM launched the current era of diffusion-based generative modelling.
Tags: generative diffusion ddpm