VGen is an open-source video generation codebase from Alibaba’s Tongyi Lab (ali-vilab) that collects several of the lab’s state-of-the-art video synthesis models in one repository. Hosted on GitHub with about 3,150 stars, it was last updated in 2025 and serves as the official implementation hub for a family of diffusion-based video models rather than a single model release.
The repository’s flagship component is I2VGen-XL, a cascaded diffusion model that turns a still image plus text into high-quality video. Around it, VGen includes VideoComposer, which adds motion controllability and compositional control over elements in the scene, and HiGen, which applies hierarchical spatio-temporal decoupling to text-to-video generation for better consistency. Training and inference code, checkpoints and configuration recipes are provided for each method, making the repo a practical starting point for reproducing and extending Alibaba’s published video research.
I2VGen-XL itself works in two cascaded stages: a base diffusion model first produces a semantically coherent low-resolution video from the input image and text, and a refinement model then raises resolution and detail while preserving motion. VideoComposer takes a different angle, conditioning generation on signals such as depth maps, sketches and motion vectors so creators can steer structure and movement explicitly. Together the included methods cover the main control axes of modern video synthesis, which is why VGen is often used as a baseline suite in later video-generation papers.
VGen is free, open-source research code published on GitHub. There is no hosted service or paid tier; you run the models on your own GPU hardware.
Video generation researchers and engineers who want production-quality reference implementations of leading diffusion video models to study, fine-tune or build upon.