We introduce Beta Diffusion, a non-Gaussian diffusion model whose forward and reverse processes are defined on bounded supports through Beta distributions. The resulting framework is naturally suited to bounded data such as pixel intensities and probability vectors, and produces sharp, high-quality samples while keeping the simplex constraint exact.
Recommended citation: M. Zhou, T. Chen, H. Zheng, and Z. Wang. Beta Diffusion. NeurIPS, 2023. · https://arxiv.org/abs/2309.07867