Accuracy and Fidelity Comparison of Luna and DALL-E 2 Diffusion-Based Image Generation Systems

Abstract

Image generation system is a system which generates images using prompts in form of text. Due to the large-scale use of automatic image generation, mainly diffusion-based systems, there needs to be an evaluation regarding the output quality generated. We conduct a qualitative analysis of the accuracy and fidelity of two image generation systems, DALL-E 2 and Luna, which differ greatly in their training datasets, algorithms, prompt handling, and output scaling. We employ a qualitative benchmarking methodology and find that DALL-E 2 outperforms Luna significantly in terms of both alignment and fidelity.