FastVideo is a framework designed to accelerate large video diffusion models, offering FastHunyuan and FastMochi for an 8x inference speedup. It provides scalable training with FSDP, sequence parallelism, and activation checkpointing, along with memory-efficient finetuning using LoRA and precomputed embeddings. The tool supports distilling, finetuning, and inferencing state-of-the-art video models like Mochi and Hunyuan. Notably, the framework achieves an 8x diffusion speed boost compared to the original Hunyuan model. FastVideo v0.1 was released on 2024/12/17 and requires Python 3.10.0, CUDA 12.1, and an H100 GPU. The framework is experimental, with ongoing development and multiple demo videos available for reference.
https://github.com/hao-ai-lab/FastVideo