FastVideo is an open-source framework from UC Berkeley’s Hao AI Lab that unifies inference and post-training for video generation models. Its goal is to make state-of-the-art video diffusion models dramatically faster to run and practical to fine-tune, without giving up output quality.
Video generation is expensive: long denoising chains make both serving and experimentation slow. FastVideo attacks both sides of the problem. On the inference side it applies sparse distillation techniques that cut denoising cost by more than 50x, backed by scalable infrastructure such as FSDP2 and sequence parallelism and multiple attention backends. On the training side it provides full fine-tuning and LoRA recipes so teams can adapt video models to their own data and style. A command-line interface and Python API cover the main workflows, and the framework runs on Linux, Windows and macOS, including consumer-grade GPUs, which puts serious video-generation research within reach of smaller labs. The project also showcases Dreamverse, a real-time generation and editing platform built on top of it. Its sparse-attention designs target both image-to-video and text-to-video workloads, and the project publishes its recipes so results can be reproduced and extended by other labs. The codebase is organized so new video models can be onboarded by adding configurations rather than rewriting the serving stack. FastVideo is Apache-2.0 licensed, has more than 4,600 stars on GitHub, and was under active development in 2026.
Free. FastVideo is open source under the Apache-2.0 license and runs entirely on your own hardware.
Researchers and engineering teams working on video generation who need faster inference, reproducible post-training pipelines, and a single codebase that spans experimentation, fine-tuning and deployment.