🎥 Frustrated by Sora's credit limits? Still waiting for Veo 2? 🚀 Open-source video DiTs are actually on par. We introduce FastVideo, an open-source stack to support fast video generation for SoTA open models. We have supported Mochi and Hunyuan, 8x faster inference, 720P 5-second video in 62 seconds.

(2/6) 🎬Compared to the original Hunyuan Video, FastVideo reduces the diffusion time from 232 seconds to 27 seconds, and the end-to-end time from 267 seconds to 62 seconds. (all measured on 8XH100)

(3/6) 🎥 Compared to the original Mochi, FastMochi reduces the diffusion time from 63 seconds to 26 seconds, and the end-to-end time from 123 seconds to 81 seconds. (all measured on 8XH100)

(4/6) Behind the scenes, FastVidep uses consistency distillation (CD). CD was proposed to accelerate image diffusion models, but its application to video Diffusion Transformers (DiT) has been scattered—until now. We’re sharing the first open recipe for CD on video DiTs with open data, checkpoints, and codebase. You can follow our recipe to distill your own model! 🚀 HF link:

(5/6) Beyond CD, FastVideo is lightweight yet powerful, packed with many useful features: 🏆 Support for distilling, finetuning, and inferencing SoTA video DiTs: Mochi and Hunyuan ⚡ Scalable training with FSDP, sequence parallelism, and selective activation checkpointing—achieving near-linear scaling to 64 GPUs. 🛠️ Memory-efficient fine-tuning with LoRA.

(6/6) Github: Shout out to @PY_Z001 @RunlongSu @_foreverpiano @BrianChen112900 @DachengLi177 and @haozhang for their hard work! We thank @anyscalecompute @richliaw and @mbzuai for their support throughout this project.

Creators, receiving payments shouldn't be a hassle. With destream, it's fast and seamless. Boost your income today!

but can you add image to video, instead of just text to video?

How many steps?

FastMochi-8 FastHunayuan-6

Firstly thanks that really cool ! Been trying the model since this afternoon but can’t figure proper setting they often come « overcooked » :| (tried the 6 steps / 6 guidance 17 flow shift)

@DrYangSong What are the min VRAM/GPU requirements for inference?