COVE: Unleashing the Diffusion Feature Correspondence for Consistent Video Editing

Jiangshan Wang1* Yue Ma1* Jiayi Guo1* Yicheng Xiao1 Gao Huang1✝ Xiu Li1✝
1Tsinghua University
Arxiv Preprint.
*Contribute Equally; ✝Corresponding Author
Paper Code (Coming soon) arXiv

Abstract

Video editing is an emerging task, in which most current methods adopt the pre-trained text-to-image (T2I) diffusion model to edit the source video in a zero-shot manner. Despite extensive efforts, maintaining the temporal consistency of edited videos remains challenging due to the lack of temporal constraints in the regular T2I diffusion model. To address this issue, we propose COrrespondence-guided Video Editing (COVE), leveraging the inherent diffusion feature correspondence to achieve high-quality and consistent video editing. Specifically, we propose an efficient sliding-window-based strategy to calculate the similarity among tokens in the diffusion features of source videos, identifying the tokens with high correspondence across frames. During the inversion and denoising process, we sample the tokens in noisy latent based on the correspondence and then perform self-attention within them. To save GPU memory usage and accelerate the editing process, we further introduce the temporal-dimensional token merging strategy, which can effectively reduce redundancy. COVE can be seamlessly integrated into the pre-trained T2I diffusion model without the need for extra training or optimization. Extensive experiment results demonstrate that COVE achieves the start-of-the-art performance in various video editing scenarios, outperforming existing methods both quantitatively and qualitatively.

Method Overview

model_overview

We propose the efficient sliding-window-based strategy for calculating the correspondence relationship among tokens in the diffusion feature of the source video. During the inversion and denoising process, the tokens are sampled based on their correspondence relationship. Through the self-attention among the corresponded tokens across frames, the quality and temporal consistency of edited videos are both significantly enhanced.

Correspondence Calculation

model_overview

Experiment Results

COVE can effectively handle various types of prompts, generating high-quality videos. For both global editing (e.g., style transferring and background editing) and local editing (e.g., modifying the appearance of the subject), COVE demonstrates outstanding performance.

(We are still refining this homepage. More experiment results will be added in the future.)

BibTeX

 @article{wang2024cove,
        title={COVE: Unleashing the Diffusion Feature Correspondence for Consistent Video Editing},
        author={Wang, Jiangshan and Ma, Yue and Guo, Jiayi and Xiao, Yicheng and Huang, Gao and Li, Xiu},
        journal={arXiv preprint arXiv:2406.08850},
        year={2024}
      }