Introducing 'FastVideo' that can speed up existing video generation AI by 8 times

A framework called ' FastVideo ' has been developed that can speed up the generation speed of video generation AI. It has been confirmed that this technology can speed up the generation speed of existing video generation models by 8 times.

GitHub - hao-ai-lab/FastVideo: FastVideo is an open-source framework for accelerating large video diffusion model.
https://github.com/hao-ai-lab/FastVideo

FastVideo is a framework developed by the Hao AI Lab, an AI lab at the University of California, San Diego, which can be used to distill existing video generation models to increase the generation speed.

Below is a video comparing the video generation speed of the open source video generation model ' HunyuanVideo ' (left) and the FastVideo applied version 'FastHunyuan' (right). In a system equipped with eight H100s, FastHunyuan can generate videos eight times faster than HunyuanVideo.

Demo of 'FastVideo' that can speed up video generation AI [FastHunyuan] - YouTube


In addition, the video below compares the generation speed of the video generation model ' Mochi ' (left) and the FastVideo version 'FastMochi'. FastMochi can also generate videos 8 times faster than Mochi.

Demo of 'FastVideo' that can speed up video generation AI [FastMochi] - YouTube


Below is a video comparing the quality of videos generated by ' Sora ', 'HunyuanVideo' and 'FastHunyuan'. It can be seen that FastVideo can speed up the video generation model without reducing its quality.

Quality comparison of high-speed video generation AI model 'HunyuanVideo' and OpenAI 'Sora' - YouTube


FastVideo is open source and can be run on GPUs with 80GB of VRAM. FastHunyuan and FastMochi are available at the following links:

FastVideo/FastHunyuan · Hugging Face
https://huggingface.co/FastVideo/FastHunyuan

FastVideo/FastMochi-diffusers · Hugging Face
https://huggingface.co/FastVideo/FastMochi-diffusers