EchoVideo: Identity-Preserving Human Video Generation by Multimodal Feature Fusion

This repo contains PyTorch model definitions, pre-trained weights and inference code for our video generation model, EchoVideo.

EchoVideo: Identity-Preserving Human Video Generation by Multimodal Feature Fusion

News

[2025.02.27] We release the inference code and model weights of EchoVideo.

Introduction

EchoVideo is capable of generating a personalized video from a single photo and a text description. It excels in addressing issues related to "semantic conflict" and "copy-paste" problems. And demonstrates state-of-the-art performance.

Gallery

1. Text-to-Video Generation

Face-ID Preserving	Full-Body Preserving

2. Comparisons

EchoVideo	ConsisID	IDAnimator

Usage

Python version is between 3.10 and 3.12, inclusive of both 3.10 and 3.12. Support both gpu and npu

cloning the repository:

git clone https://github.com/bytedance/EchoVideo
cd EchoVideo

Installation

pip install -r requirements.txt

Download Pretrained Weights

The details of download pretrained models are shown here.

Run Demo

# multi-resolution video generation [(480, 640), (480, 848), (480, 480), (848, 480), (640, 480)]
python infer.py

Methods

Overall Architecture

Overall architecture of EchoVideo. By employing a meticulously designed IITF module and mitigating the over-reliance on input images, our model effectively unifies the semantic information between the input facial image and the textual prompt. This integration enables the generation of consistent characters with multi-view facial coherence, ensuring that the synthesized outputs maintain both visual and semantic fidelity across diverse perspectives.

Key Features

Illustration of facial information injection methods. (a) IITF. Facial and textual information are fused to ensure consistent guidance throughout the generation process. we propose IITF to fuse text and facial information, establishing a semantic bridge between facial and textual information, coordinating the influence of different information on character features, thereby ensuring the consistency of generated characters. IITF consists of two core components: facial feature alignment and conditional feature alignment. (b) Dual branch. Facial and textual information are independently injected through Cross Attention mechanisms, providing separate guidance for the generation process.

Benchmark

Model	Identity Average↑	Identity Variation↓	Inception Distance↓	Dynamic Degree↑
IDAnimator	0.349	0.032	159.11	0.280
ConsisID	0.414	0.094	200.40	0.871
pika	0.329	0.091	268.35	0.954
Ours	0.516	0.075	176.53	0.955

Acknowledgements

CogVideo: The DiT module we adpated from, and the VAE module we used. MODEL LICENSE
SigLip: Vision Encoder we used.

BibTeX

If you find our work useful in your research, please consider citing the paper

@article{wei2025echovideo,
  title={EchoVideo: Identity-Preserving Human Video Generation by Multimodal Feature Fusion},
  author={Wei, Jiangchuan and Yan, Shiyue and Lin, Wenfeng and Liu, Boyuan and Chen, Renjie and Guo, Mingyu},
  journal={arXiv preprint arXiv:2501.13452},
  year={2025}
}