Text-to-image (T2I) diffusion models achieve state-of-the-art results in image synthesis and editing. However, leveraging such pre-trained models for video editing is considered a major challenge. Many existing works attempt to enforce temporal consistency in the edited video through explicit correspondence mechanisms, either in pixel space or between deep features. These methods, however, struggle with strong nonrigid motion. In this paper, we introduce a fundamentally different approach, which is based on the observation that spatiotemporal slices of natural videos exhibit similar characteristics to natural images. Thus, the same T2I diffusion model that is normally used only as a prior on video frames, can also serve as a strong prior for enhancing temporal consistency by applying it on spatiotemporal slices. Based on this observation, we present Slicedit^#, a method for text-based video editing that utilizes a pre-trained T2I diffusion model to process both spatial and spatiotemporal slices. Our method generates videos that retain the structure and motion of the original video while adhering to the target text. Through extensive experiments, we demonstrate Slicedit's ability to edit a wide range of real-world videos, confirming its clear advantages compared to existing baselines.

^#Slicedit can be pronounced ``slice-edit'' or ``sliced-it''

To edit a video, I₀, we apply DDPM inversion using our video-denoising model, which is an inflated version of the T2I model. This process extracts noise volumes and attention maps for each diffusion timestep. Subsequently, we run DDPM sampling using the extracted noise space, while injecting the extended attention maps at specific timesteps, as can be seen on the left pane of the figure. The inversion and sampling are performed while conditioning the inflated denoiser on the source and target text prompts, respectively.
Our inflated denoiser can be seen on the right pane. It employs two versions of the pre-trained image denoiser. A version with extended attention is applied to x - y slices (blue), and the original denoiser is applied to y - t slices (red). The two predicted noise volumes are then combined into the final predicted noise volume (marked in green).

See our paper for more details.

Paper

Slicedit: Zero-Shot Video Editing With Text-to-Image Diffusion Models Using Spatio-Temporal Slices
Nathaniel Cohen^*, Vladimir Kulikov^*, Matan Kleiner^*, Inbar Huberman-Spiegelglas, Tomer Michaeli.
Arxiv Paper Supplementary

Bibtex

@article{
	cohen2024slicedit,
	title={Slicedit: Zero-Shot Video Editing With Text-to-Image Diffusion Models Using Spatio-Temporal Slices},
	author={Cohen, Nathaniel and Kulikov, Vladimir and Kleiner, Matan and Huberman-Spiegelglas, Inbar and Michaeli, Tomer},
	journal={arXiv preprint arXiv:2405.12211},
	year={2024},
}
		

Code

References

[1] Geyer, Michal, Omer Bar-Tal, Shai Bagon, and Tali Dekel. "Tokenflow: Consistent diffusion features for consistent video editing." ICLR 2024.
[2] Ceylan, Duygu, Chun-Hao P. Huang, and Niloy J. Mitra. "Pix2video: Video editing using image diffusion." ICCV 2023.
[3] Zhang, Yabo, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. "Controlvideo: Training-free controllable text-to-video generation." ICLR 2024.
[4] Yang, Shuai, Yifan Zhou, Ziwei Liu, and Chen Change Loy. "Rerender A Video: Zero-Shot Text-Guided Video-to-Video Translation." SIGGRAPH Asia 2023.
[5] Wu, Jay Zhangjie, 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." ICCV 2023.
[6] Cong, Yuren, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. "FLATTEN: optical FLow-guided ATTENtion for consistent text-to-video editing." ICLR 2024.