UCSD's Hao AI Lab releases FastWan, generating 5-second 480P video in 5 seconds via sparse distillation
AI video model releases
UCSD's Hao AI Lab releases FastWan, generating 5-second 480P video in 5 seconds via sparse distillation
FastWan2.1-1.3B generates a 5-second 480P video end-to-end in 5 seconds on a single H200 GPU (denoising takes only 1 second), and in 21 seconds on a single RTX 4090.
FastWan2.2-5B generates a 5-second 720P video in 16 seconds on a single H200.
The core method, called sparse distillation, is the first to jointly train sparse attention and denoising-step distillation together, cutting Wan2.1-14B's 50 diffusion steps down to as few as 3 while retaining sparse-attention speedups.
Using VSA + DMD + torch compile, Wan2.1-14B 720P denoising time drops from 1746.5s (FA2 baseline) to 13s, and Wan2.1-1.3B 480P drops from 95.21s to 0.98s.
All model weights, training recipes, and datasets (including FastWan2.1-T2V-1.3B and FastWan2.2-TI2V-5B) are released under the Apache-2.0 license on GitHub and Hugging Face, with a 14B sparse-distilled version still in progress.
cho.sh