English

FSVideo: Fast Speed Video Diffusion Model in a Highly-Compressed Latent Space

Computer Vision and Pattern Recognition 2026-02-03 v1

Abstract

We introduce FSVideo, a fast speed transformer-based image-to-video (I2V) diffusion framework. We build our framework on the following key components: 1.) a new video autoencoder with highly-compressed latent space (64×64×464\times64\times4 spatial-temporal downsampling ratio), achieving competitive reconstruction quality; 2.) a diffusion transformer (DIT) architecture with a new layer memory design to enhance inter-layer information flow and context reuse within DIT, and 3.) a multi-resolution generation strategy via a few-step DIT upsampler to increase video fidelity. Our final model, which contains a 14B DIT base model and a 14B DIT upsampler, achieves competitive performance against other popular open-source models, while being an order of magnitude faster. We discuss our model design as well as training strategies in this report.

Keywords

Cite

@article{arxiv.2602.02092,
  title  = {FSVideo: Fast Speed Video Diffusion Model in a Highly-Compressed Latent Space},
  author = {FSVideo Team and Qingyu Chen and Zhiyuan Fang and Haibin Huang and Xinwei Huang and Tong Jin and Minxuan Lin and Bo Liu and Celong Liu and Chongyang Ma and Xing Mei and Xiaohui Shen and Yaojie Shen and Fuwen Tan and Angtian Wang and Xiao Yang and Yiding Yang and Jiamin Yuan and Lingxi Zhang and Yuxin Zhang},
  journal= {arXiv preprint arXiv:2602.02092},
  year   = {2026}
}

Comments

Project Page: https://kingofprank.github.io/fsvideo/

R2 v1 2026-07-01T09:31:49.782Z