Improved Zero-Shot Image Editing via Null-Toon and Directed Delta Denoising Score

Abstract

Recently, there has been a rapid surge in the utilization of diffusion models for customized image generation and editing tasks, especially using zero-shot editing algorithms that can largely operate on given images regardless of their source domain. This work is based on two well-known zero-shot image editing algorithms: Null Text Inversion (NTI) and Delta Denoising Score (DDS). With respect to NTI, we mainly focus on image cartoonization, which has received less attention in the context of text-guided image editing. In a nutshell, we propose a customized reconstruction phase for NTI, which helps transforming the natural input image into cartoon images with desired customization by supporting parameters. We also improve the current DDS optimization baseline and propose the Directed Delta Denoising Score (DDDS). Our DDDS algorithm offers a better image editing experience by replacing the target text prompt with the proposed directed text prompt. Computing directed text prompt requires one subtraction operation and yields significant reconstruction improvement over DDS. To demonstrate the effectiveness of our contributions, the paper presents both quantitative and qualitative comparisons against the state-of-the-art, as well as several visual examples.