GitHub - hao-ai-lab/FastVideo: A unified inference and post-training framework for accelerated video generation. · GitHub

Generate 5 seconds of video in just 1.8 seconds end-to-end on a single GPU. FastVideo is a highly optimized, unified post-training and real-time inference framework designed for accelerated video generation, pushing the boundary of open-source video diffusion transformers.

🔹 **Real-Time Video Generation**: Utilizing the newly released FastWan-QAD model, the framework runs a complete end-to-end generation of a 5-second video in only 1.8 seconds, delivering faster-than-real-time rendering capabilities.

🔹 **Sparse Distillation Technology**: Integrates proprietary Video Sparse Attention alongside step-by-step sparse distillation techniques. This achieves a massive 50x denoiser speedup, allowing high-quality video generation on consumer-grade hardware.

🔹 **Unified Training Pipeline**: Supports both full parameter finetuning and LoRA adapters for state-of-the-art open DiT architectures, offering developers a complete data preprocessing pipeline for text, image, and video content.

🔹 **Scalable Distributed Performance**: Implements advanced memory-saving techniques including FSDP2, sequence parallelism, and selective activation checkpointing. These scale multi-GPU distributed training and inference seamlessly without bottlenecking.

🔹 **Advanced Denoising Algorithms**: Adapts Distribution Matching Distillation (DMD2) and causal distillation through Self-Forcing, giving developers granular control over video coherence, motion stability, and generation quality.

https://github.com/hao-ai-lab/FastVideo