Announcing FastVideo V1, a unified framework for accelerating video generation. FastVideo V1 offers: - A simple, consistent Python API - State of the art model performance optimizations - Optimized implementations of popular models Blog:

FastVideo V1 integrates SageAttention and Teacache, delivering 3x generation speedup and 7x model loading speedup while maintaining quality. Our design allows for future models and optimizations to be seamlessly added and composed together.

FastVideo supports multi-GPU inference and optimizations for state-of-the-art video diffusion models, with more to come soon! See the current support matrix below:

Never wrestle with torchrun commands or messy bash scripts for multi-GPU generation again! FastVideo V1 uses optimal default parameters while offering full configurability for out-of-the-box multi-GPU inference. If you're a Diffusers user frustrated by 15+ minute generation times and complex multi-GPU setups, FastVideo provides a simpler alternative for state-of-the-art models like Wan2.1.

We are grateful to the following projects we learned and reused code from: PCM, Diffusers, OpenSoraPlan, xDiT, vLLM, SGLang, Wan2.1. FastVideo V1 development was supported by @anyscalecompute and @mbzuai . We also appreciate our early testers and community, especially @jd_markovchain and @yusufozuysal , for valuable feedback. FastVideo Team: @wlsaidhi, @PY_Z001, @WZhou35897, @KevinLin61672, @BrianChen112900, Zihang He, You Zhou, Wenting Zhang, @CodyHaoYu, @richliaw, @haozhangml