A Video is Worth 256 Bases: Spatial-Temporal Expectation-Maximization Inversion for Zero-Shot Video Editing

Maomao Li1,2, Yu Li2, Tianyu Yang2, Yunfei Liu2, Dongxu Yue3, Zhihui Lin4, Dong Xu1
1The University of Hong Kong    2International Digital Economy Academy (IDEA)
3Peking University    4Tsinghua University
Code    arXiv

We propose STEM inversion as an alternative approach to zero-shot video editing, which offers several advantages over the commonly employed DDIM inversion technique. STEM inversion achieves superior temporal consistency in video reconstruction while preserving intricate details. Moreover, it seamlessly integrates with contemporary video editing methods, such as TokenFlow [2] and FateZero [1], enhancing their editing capabilities. We also provide qualitative comparison between different video editing methods: Tune-A-Video [3], Pix2Video [4], Text2Video-Zero [5]. We recommend using 1080P to watch video editing results.

Abstract

This paper presents a video inversion approach for zero-shot video editing, which aims to model the input video with low-rank representation during the inversion process. The existing video editing methods usually apply the typical 2D DDIM inversion or naive spatial-temporal DDIM inversion before editing, which leverages time-varying representation for each frame to derive noisy latent. Unlike most existing approaches, we propose a Spatial-Temporal Expectation-Maximization (STEM) inversion, which formulates the dense video feature under an expectation-maximization manner and iteratively estimates a more compact basis set to represent the whole video. Each frame applies the fixed and global representation for inversion, which is more friendly for temporal consistency during reconstruction and editing. Extensive qualitative and quantitative experiments demonstrate that our STEM inversion can achieve consistent improvement on two state-of-the-art video editing methods.

Method

We argue that frames over a larger range should be considered to execute DDIM inversion. However, using all video frames directly will bring unacceptable complexity. To deal with this, we propose a Spatial-Temporal Expectation-Maximization (STEM) inversion method. The insight behind this is massive intra-frame and inter-frame redundancy lie in a video, thus there is no need to treat every pixel in the video as reconstruction bases. Then, we use the EM algorithm to find a more compact basis set (e.g., 256 bases) for the input video.

MY ALT TEXT

The illustration of the proposed STEM inversion. We estimate a more compact representation (bases $\mu$) for the input video via the EM algorithm. The ST-E step and ST-M step are executed alternately for R times until convergence. The Self-attention (SA) in our STEM inversion are denoted as STEM-SA, where the $\rm{Key}$ and $\rm{Value}$ embeddings are derived by projections of the converged $\mu$.

Reconstruction Comparison

Feature visualization

MY ALT TEXT

Left: we first estimate the optical flow of the input sequence. Then, we apply PCA on the output features of the last SA layer from the UNet decoder. The 4-th column shows the feature visualization when we use optical flow to warp the former-frame features. Last, we give the cosine similarity of the warped features and the target ones. Here, the brighter, the better. Right: we provide the mean cosine similarity across different time steps. The higher similarity indicates that our STEM inversion can achieve better temporal consistency from the perspective of optical flow.

BibTeX


        @article{stem-inv,
        title = {A Video is Worth 256 Bases: Spatial-Temporal Expectation-Maximization Inversion for Zero-Shot Video Editing},
        author = {Maomao Li, Yu Li, Tianyu Yang, Yunfei Liu, Dongxu Yue, Zhihui Lin, and Dong Xu},
        journal={arXiv preprint arxiv:2312.05856},
        year={2023}
        }

Reference

[1] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. ICCV, 2023.
[2] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing.
[3] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. CVPR, 2023.
[4] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. ICCV, 2023.
[5] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. ICCV, 2023