FastVideo
A unified post-training and real-time inference framework for accelerated video generation, supporting full and LoRA fine-tuning, sparse attention and distillation techniques that cut denoising steps and GPU time.
FastVideo is a post-training and inference framework for video-generation models, supporting full and LoRA fine-tuning of open video diffusion transformers, a data-preprocessing pipeline for video, image and text, and distillation techniques including Distribution Matching Distillation and causal distillation through self-forcing. Its Video Sparse Attention and sparse-distillation methods report large denoising speedups on supported models.
Training scales with FSDP2, sequence parallelism and selective activation checkpointing, and released checkpoints and recipes, including the FastWan and FastH3 model families, run locally on Apple Silicon through MLX and on NVIDIA hardware including DGX Spark through CUDA. The project is developed with partners including Nuva Lab and the NVIDIA FastGen team.