VideoCanvas: Unified Video Completion from Arbitrary Spatiotemporal Patches via In-Context Conditioning

Minghong Cai^{1 †}, Qiulin Wang^{2 ✉}, Zongli Ye¹, Wenze Liu¹, Quande Liu², Weicai Ye², Xintao Wang², Pengfei Wan², Kun Gai², Xiangyu Yue^{1 ✉
1}MMLab, The Chinese University of Hong Kong ²Kling Team, Kuaishou Technology
†: Intern at Kuaishou Technology, ✉: Corresponding Authors

📋 News

• [2025.10.9] Release Arxiv paper.

📖 Introduction

VideoCanvas has two key contributions:

• 🎯 Unified Tasks: VideoCanvas introduces a unified paradigm for arbitrary spatio-temporal video generation, seamlessly integrating diverse capabilities including image/patch-to-video conditioning at any timestamp, inpainting/outpainting, camera control, scene transitions, and video extension.
• 🛠️ Simple Solution: Our technical innovation leverages In-Context Conditioning with zero-padding for spatial control and Temporal RoPE Interpolation for temporal alignment, achieving frame-precise video generation without fine-tuning VAEs or adding parameters.
  

https://github.com/user-attachments/assets/c3777c16-0a1e-4cbb-af6f-856f88312317

📖 VideoCanvasBench

We will release this benchmark, including intra-scene and inter-scene evaluation data.

⚙️ Code (Coming soon)

Citation

 @article{cai2025videocanvas,
    title={VideoCanvas: Unified Video Completion from Arbitrary Spatiotemporal Patches via In-Context Conditioning},
    author={Minghong Cai, Qiulin Wang, Zongli Ye, Wenze Liu, Quande Liu, Weicai Ye, Xintao Wang, Pengfei Wan, Kun Gai, Xiangyu Yue},
    journal={arXiv preprint arXiv:2510.08555},
    year={2025}
}