VideoSys

Why we included this project

VideoSys is a system for running text-to-video diffusion models without wiring up each upstream codebase yourself. It puts Open-Sora, CogVideoX, Latte, Vchitect, and Open-Sora-Plan behind a single training and inference interface, so swapping between generators or comparing them in one pipeline is straightforward. Where it stands out is speed: the project contributes published techniques like Dynamic Sequence Parallelism for multi-GPU training, Data-Centric Parallel for batches of variable-length videos, and Pyramid Attention Broadcast (PAB), a training-free method that reaches real-time DiT-based generation with negligible quality loss. If you are building a video generation service or fine-tuning these base models on your own data, this gives you a faster, more memory-efficient starting point than assembling each upstream project separately. It is a system to adopt and run, not a model to consume directly.