FastWan Generating a 5-Second Video in 5 Seconds via Sparse Distillation

Gradio app: fastwan.fastvideo.org

Video generation can be accelerated even further with sparse attention!

FastWan introduces a new training method called Sparse Distillation, applied to DiT-based video diffusion models, that combines sparse attention and diffusion step reduction to enable real-time video generation. It demonstrates that sparse attention and step compression can be

Nice -- try FastWan at fal !

🎉 Congratulations to the Hao AI Lab on enabling users to produce a video in just 5 Seconds with FastVideo. This is done using FastWan, a new family of video generation models that were trained with Sparse Distillation to speed up video denoising time by 70X. 📹 5-second 480p

Thank you NVIDIA Data Center for featuring our recent results!☺️😘

D2F Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing

This is a very exciting work I recently get involved. TL;DRs: - while diffusion LLM sounds fancy, in fact it decodes *way slower* than autoregressive decoding. This is true for almost all OSS dLLMs - this work develops a set of very practical optimizations to make dLLM run faster

⚡ Turn ideas into videos in seconds — with fastWan by Hao AI Lab, powered by GMI Cloud. ✨ Fast. Fun. Full of imagination. 🚀 No waiting. No limits. Just pure creativity. 🎥 HaoAILabs × GMI Cloud — speed meets charm. Watch the magic happen 👇 fastwan.fastvideo.org #fastWan