How to Improve Performance of a Diffusion Model by an Ensemble of Expert Models

A comprehensive explanation and review of the ensemble-based diffusion model from NVIDIA

In this article we shall go through a recent development in the Diffusion Model domain called eDiff-I [1]. It is an ensemble of Diffusion Models which seems to have outperformed all the other state-of-the-art (e.g., DALL-E2 and Stable Diffusion Model). This article will build upon earlier concepts (e.g., Diffusion Models and Stable Diffusion Models). Feel free to checkout my earlier posts on these topics.

Figure 1: Binary Tree Diagram Depicting an Ensemble of Diffusion Models being trained at Various Noise Levels (Source: Author)