CustomVideo: Customizing Text-to-Video Generation with Multiple Subjects

Zhao Wang1, Aoxue Li2, Enze Xie2, Lingting Zhu3, Yong Guo2, Qi Dou1, Zhenguo Li2
1The Chinese University of Hong Kong   2Huawei Noah Ark's Lab   3The University of Hong Kong

arXiv

Customized text-to-video generation results of our proposed CustomVideo with given multiple subjects and text prompts. Our approach can disentangle highly similar subjects, e.g., cat v.s. dog, preserving the fidelity of subjects and smooth motions.


Abstract

Customized text-to-video generation aims to generate high-quality videos guided by text prompts and subject references. Current approaches designed for single subjects suffer from tackling multiple subjects, which is a more challenging and practical scenario. In this work, we aim to promote multi-subject guided text-to-video customization. We propose CustomVideo, a novel framework that can generate identity-preserving videos with the guidance of multiple subjects. To be specific, firstly, we encourage the co-occurrence of multiple subjects via composing them in a single image. Further, upon a basic text-to-video diffusion model, we design a simple yet effective attention control strategy to disentangle different subjects in the latent space of diffusion model. Moreover, to help the model focus on the specific object area, we segment the object from given reference images and provide a corresponding object mask for attention learning. Also, we collect a multi-subject text-to-video generation dataset as a comprehensive benchmark, with 69 individual subjects and 57 meaningful pairs. Extensive qualitative, quantitative, and user study results demonstrate the superiority of our method, compared with the previous state-of-the-art approaches.

Method

We propose a simple yet effective co-occurrence and attention control mechanism with mask guidance to preserve the the fidelity of subjects for multi-subject driven text-to-video generation. During the training stage, only the key and value weights in the cross attention layers are fine-tuned. In the inference stage, given a text prompt integrated with learned text token, we can easily obtain high-quality videos with specific subjects.

Comparisons

Subject1

Subject2

      CustomVideo

         DreamBooth [1]

      CustomDiffusion [2]

      VideoDreamer [3]

Qualitative results of our CustomVideo with comparison to SOTA methods. We can observe that our CustomVideo
can generate videoswith much better fidelity of subjects compared with previous SOTA methods.

BibTeX


@article{wang2024customvideo,
  title={CustomVideo: Customizing Text-to-Video Generation with Multiple Subjects},
  author={Wang, Zhao and Li, Aoxue and Xie, Enze and Zhu, Lingting and Guo, Yong and Dou, Qi and Li, Zhenguo},
  journal={arXiv preprint arXiv:2401.09962},
  year={2024}
}

[1] Ruiz, Nataniel, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. "Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation." CVPR, 2023.

[2] Kumari, Nupur, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. "Multi-concept customization of text-to-image diffusion." CVPR, 2023.

[3] Chen, Hong, Xin Wang, Guanning Zeng, Yipeng Zhang, Yuwei Zhou, Feilin Han, and Wenwu Zhu. "Videodreamer: Customized multi-subject text-to-video generation with disen-mix finetuning." arXiv preprint arXiv:2311.00990, 2023.