FastVideo

Speeding up video diffusion is FastVideo's whole agenda: the UCSD Hao AI Lab framework unifies post-training and real-time inference for video generation models, wrapping techniques like Video Sparse Attention, sparse distillation, and self-forcing causal distillation into one pipeline. Its FastWan models, including FastWan2.1-T2V-1.3B and FastWan2.2-TI2V-5B, apply distillation to the open Wan models to reach faster-than-real-time generation of short clips, and the framework also supports full fine-tuning, LoRA fine-tuning, and distribution matching distillation for other state-of-the-art open video DiTs. Sequence parallelism spreads inference across GPUs, with H100, A100, and RTX 4090 hardware supported on Linux, Windows, and macOS, including an MPS path for Apple Silicon. Installation uses the uv package manager with a CUDA-matched PyTorch backend. Everything is Apache-2.0 with about 3,900 GitHub stars, and documentation lives at hao-ai-lab.github.io/FastVideo. Researchers and builders use it to make heavy open video models cheap enough for interactive use.