VideoSys is an open-source system from NUS HPC-AI Lab that makes video generation with diffusion models both easy to use and fast to run. Where most text-to-video repositories ship a single model with a research-grade script, VideoSys provides a unified pipeline abstraction across multiple video models and layers on acceleration techniques such as its PAB (Pyramid Attention Broadcast) caching, so the same prompts generate in a fraction of the time. The project has around 2,000 GitHub stars and was last updated in 2025.
VideoSys exposes a simple generate() style API: pick a supported model (the framework covers popular open video diffusion models such as Open-Sora, CogVideoX and Latte-class architectures), set resolution and steps, and run. Under the hood it applies training-free acceleration: PAB reuses attention outputs across timesteps where they barely change, and the engine supports distributed inference strategies for multi-GPU setups. Benchmarks in the README show substantial speed-ups with negligible quality loss, which matters because video diffusion is one of the most compute-hungry generative workloads.
Free. VideoSys is open source under the Apache-2.0 license; you supply the GPU compute for generation.
Researchers and engineers working with open video-generation models who need faster inference and a single framework instead of a pile of incompatible per-model scripts.
VideoSys comes from the same research group behind several efficient-diffusion projects, and its engineering focus shows in the details: memory-efficient attention, timestep caching policies and sequence parallelism are exposed as composable options rather than hard-coded hacks. For practitioners, that means a new video model can usually be onboarded by writing a thin pipeline wrapper, immediately inheriting the acceleration work instead of re-implementing it for every release.