VidToMe: Video Token Merging for Zero-Shot Video Editing

CVPR 2024
Xirui Li1, Chao Ma1, Xiaokang Yang1, Ming-Hsuan Yang2
1Shanghai Jiao Tong University, 2UC Merced
Paper Code arXiv Video

Abstract

Diffusion models have made significant advances in generating high-quality images, but their application to video generation has remained challenging due to the complexity of temporal motion. Zero-shot video editing offers a solution by utilizing pre-trained image diffusion models to translate source videos into new ones. Nevertheless, existing methods struggle to maintain strict temporal consistency and efficient memory consumption. In this work, we propose a novel approach to enhance temporal consistency in generated videos by merging self-attention tokens across frames. By aligning and compressing temporally redundant tokens across frames, our method improves temporal coherence and reduces memory consumption in self-attention computations. The merging strategy matches and aligns tokens according to the temporal correspondence between frames, facilitating natural temporal consistency in generated video frames. To manage the complexity of video processing, we divide videos into chunks and develop intra-chunk local token merging and inter-chunk global token merging, ensuring both short-term video continuity and long-term content consistency. Our video editing approach seamlessly extends the advancements in image editing to video editing, rendering favorable results in temporal consistency over state-of-the-art methods.

Method

Given an input source video and a text prompt, our proposed VidToMe generates an edited video according to the prompt using a pretrained text-to-image diffusion model. The critical idea is to merge similar tokens across multiple frames in self-attention modules to achieve temporal consistency in generated videos.

Top: VidToMe pipeline. Starting from DDIM inverted noise latent, we denoise the video frames with the diffusion model combined with an existing controlling method. The video token merging is applied before self-attention blocks.

Bottom: Video Token Merging. Given the self-attention input tokens, we first merge tokens in the video chunk (Local Token Merging). Then, a global token set is combined to enable inter-chunk token sharing (Global Token Merging). After joint self-attention on the merged token set, the output tokens are restored to their original size by reversing the merging process.

VidToMe Editing Results

Left: Input Video; Right: Edit Video.

To "Van Gogh Style"
To "Rainbow-colored Origami Flamingo"
To "Oil Painting Style"
To "GPU in a Computer"
To "Comic Book, Black and White Pencil Sketch"
To "White Ancient Greek Sculpture"
To "Pop Art Style"
To "Volcano, Spewing Lava"

Comparisons

Efficiency

Merging multi-frame tokens reduces the temporal redundancy in video frames, improving memory and computational efficiency. Therefore, compared with baseline methods, VidToMe consumes less memory while generating videos quickly.

Using Different Control

Using Customized Models


1. Use LoRA model from artificialguybr.
2. Use customized model Anything.

Video Presentation

BibTeX

@article{li2023vidtome,
    title={VidToMe: Video Token Merging for Zero-Shot Video Editing},
    author={Li, Xirui and Ma, Chao and Yang, Xiaokang and Yang, Ming-Hsuan},
    booktitle={arXiv prepsrint arxiv:2312.10656},
    year={2023}
    }

[1] Narek Tumanyan, Michal Geyer, et al. “Plug-and-play diffusion features for text-driven image-to-image translation”. In CVPR. 2023.
[2] Robin Rombach, Andreas Blattmann, et al. “High-resolution image synthesis with latent diffusion models”. In CVPR. 2022.
[3] Levon Khachatryan, Andranik Movsisyan, et al. “Text2video-zero: Text-to-image diffusion models are zero-shot video generators”. arXiv preprint arXiv:2303.13439. 2023.
[4] Jay Zhangjie Wu, Yixiao Ge, et al. “Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation”. In ICCV, 2023.
[5] Wen Wang, kangyang Xie, et al. “Zero-shot video editing using off-the-shelf image diffusion models”. arXiv preprint arXiv:2303.17599. 2023.
[6] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. “Pix2video: Video editing using image diffusion”. In ICCV. 2023.
[7] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. “Adding conditional control to text-to-image diffusion models”. In ICCV. 2023.
[8] Edward J. Hu and Yelong Shen, et al. “LoRA: Low-Rank Adaptation of Large Language Models”. In ICLR. 2021.