VideoDreamBooth: Subject-Driven Video Editing with Guided Multi-Path Fusion

Diffusion models have gained popularity for generating content in various domains, including text, image, and video. In particular, Text-to-Image(T2I) models have emerged as an innovative tool for artistic expression. They also spurred growth in other associated areas, such as personalized subject generation and one-shot video editing. This paper presents VideoDreamBooth, a novel framework for real video editing with personalized subject. Inspired by the coarse-to-fine denoising nature of diffusion, we propose a multi-path fusion technique to preserve temporal consistency with original video. We also propose a metric \(CLIP\ score\) to guide which denoising steps will conduct fusion. Our framework has proven to be effective and successful for video editing in various experimental settings.