TokenFlow: Consistent Diffusion Features for Consistent Video Editing

Anonymous Submisssion
Supplementary Material Code

Abstract

The generative AI revolution has been recently expanded to videos. Nevertheless, current state-of-the-art video models are still lagging behind image models in terms of visual quality and user control over the generated content. In this work, we present a framework that harnesses the power of a text-to-image diffusion model for the task of text-driven video editing. Specifically, given a source video and a target text-prompt, our method generates a high-quality video that adheres to the target text, while preserving the spatial layout and dynamics of the input video. Our method is based on our key observation that consistency in the edited video can be obtained by enforcing consistency in the diffusion feature space. We achieve this by explicitly propagating diffusion features based on inter-frame correspondences, readily available in the model. Thus, our framework does not require any training or fine-tuning, and can work in conjunction with any off-the-shelf text-to-image editing method. We demonstrate state-of-the-art editing results on a variety of real-world videos.

Method

We observe that the level of temporal consistency of a video is tightly related to the temporal consistency of its feature representation, as can be seen in the feature visualization below. The features of a natural video have a shared, temporally consistent representation. When editing the video per frame, this consistency breaks. Our method ensures the same level of feature consistency as in the original video features.

Original Per Frame Editing Ours

Our key finding is that a temporally-consistent edit can be achieved by enforcing consistency on the internal diffusion features across frames during the editing process. We achieve this by propagating a small set of edited features across frames, using the correspondences between the original video features. Given an input video I, we invert each frame, extract its tokens (i.e., output features from the self-attention modules), and extract inter-frame feature correspondences using a nearest-neighbor (NN) search. At each denoising step:

    (I) We sample keyframes from the noisy video J_t and jointly edit them using an extended-attention block. The set of resulting edited tokens is T_base.
    (II) We propagate the edited tokens across the video according to the pre-computed correspondences of the original video features.
To denoise J_t, we feed each frame to the network and replace the generated tokens with the tokens obtained from the propagation step (II).

TokenFlow Editing results

Hover over the videos to see the original video and text prompts. Please see our Supplementary Materials for more results.

an ice sculpture
a robotic wolf
a marble sculpture

Van Gogh painting
an origami of a stork

a car made of ice on an icy road
Van Gogh painting

a robot spinning a silver ball
colorful crochet kittens

Comparisons

Please see our Supplementary Materials for more comparisons.

Input video

                     Ours

               Text-to-video [1]

    Tune-a-video [2]

         Gen-1 [3]

          Per frame PnP[4]

     Fate-Zero[5]

     Rerender-a-video[6]

[1] 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. arXiv preprint arXiv:2303.13439, 2023.

[2] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, 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. arXiv preprint arXiv:2212.11565, 2022

[3] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023

[4] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text- driven image-to-image translation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

[5] 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. arXiv:2303.09535, 2023 .

[6] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to- video translation, 2023.