Community Project Profile
VideoSys
An efficient system framework for video generation models
- Organisation
- NUS HPC-AI Lab
- Group
- University / research
- Category
- Video generation system
- Status
- Research open source
- Started
- 2024-02
- Language / Form
- Python
- License
- Apache-2.0
- GitHub Stars
- 2,021
- Updated
- 2026-05-04
VideoSys addresses the systems problem behind video generation: video models are compute-heavy and memory-hungry, so they need system optimization to become trainable and serviceable.
What It Is
VideoSys is a system framework for video generation, designed to make it easier to train, optimize, and run video generation models. It follows the NUS HPC-AI Lab’s systems line: not only model quality, but making models cheaper, faster, and usable.
In video generation, systems efficiency can determine whether a product can serve users at scale.
AI Relevance
Video generation is one of the costliest areas of generative AI. Text models are already expensive; video must handle time, spatial resolution, and longer contexts.
VideoSys matters because it pushes video generation from demos toward engineering systems, making deployment more realistic for research and application teams.
Singapore Relevance
VideoSys and Colossal-AI together form the AI-systems asset base of the NUS HPC-AI Lab. It shows a feasible position for Singapore in model competition: not only building the largest models, but improving the training and serving efficiency behind them.
That is especially realistic for Singapore, where compute is a visible constraint.
Milestones
- 2024-02VideoSys repository created