HyDRA: Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models

This is the official Hugging Face model repository for HyDRA (Hybrid Memory for Dynamic Video World Models).

🔗 GitHub Repository: H-EmbodVis/HyDRA 📄 Project Page: Hybrid-Memory-in-Video-World-Models

🔍 Overview

While recent video world models excel at simulating static environments, they share a critical blind spot: the physical world is dynamic. When moving subjects exit the camera's field of view and later re-emerge, current models often lose track of them.

To bridge this gap, we introduce Hybrid Memory, a novel paradigm that requires models to simultaneously act as precise archivists for static backgrounds and vigilant trackers for dynamic subjects. HyDRA is a specialized memory architecture that compresses contexts into memory tokens and utilizes a spatiotemporal relevance-driven retrieval mechanism.

🎯 Task & Capabilities

• Task: Text-to-Video Generation / Video World Modeling
• Input: Text prompts, camera poses, and initial video latents.
• Output: High-fidelity video sequences maintaining both identity and motion continuity of dynamic subjects, even during out-of-view intervals.

🚀 Usage

To use these weights, please refer to our GitHub repository: H-EmbodVis/HyDRA

📖 Citation

If you find our work useful, please consider citing:

@article{chen2026out,
  title   = {Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models},
  author  = {Chen, Kaijin and Liang, Dingkang and Zhou, Xin and Ding, Yikang and Liu, Xiaoqiang and Wan, Pengfei and Bai, Xiang},
  journal = {arXiv preprint arXiv:2603.25716},
  year    = {2026}
}