FastVideo Open-Source Framework Delivers 50x Video Generation Speedup Across NVIDIA, Apple Silicon, and Cloud Platforms

FastVideo, a new open-source framework, delivers a groundbreaking 50x video generation speedup across NVIDIA GPUs, Apple Silicon, and cloud platforms, featuring sparse distillation, LoRA finetuning, and real-time inference capabilities, with recent updates adding synchronized video-audio generation and Apple Silicon-optimized models.

• FastVideo is an open-source unified framework for accelerated video generation, supporting both post-training workflows and real-time inference with features like sparse distillation, LoRA finetuning, and Distribution Matching Distillation achieving over 50x denoising speedup.
• The framework supports diverse hardware including NVIDIA GPUs, Apple Silicon via MLX, and NVIDIA DGX Spark, and runs on Linux, Windows, and macOS, with a user-friendly Python API and CLI for generating videos in just a few lines of code.
• Recent updates include FastH3 Preview v0.2 for synchronized video and audio generation, FastMetal-QAD models optimized for Apple Silicon, and the Dreamverse platform enabling real-time 'vibe directing' video streaming deployable locally or via cloud services.