BREAKING
NVIDIA Speeds Up Video Diffusion 5-10x
0
x
Peak speedup
0
x
Min speedup
Two Techniques Cut the Load
1
Autoregressive frames
↓
2
Temporal cache compression
↓
3
Sparse attention
↓
4
Fast long rollout
Old Diffusion vs New Method
Bidirectional
Slow
●
Too slow for interactive use
●
Growing memory load
Autoregressive
New
●
Frame by frame
●
Constant memory
NVIDIA's Footprint at ICML 2026
Accepted papers
74
Cite NVIDIA GPUs
2000
Cite Nemotron
145
Still Research Stage
AI NEWS BLITZ
At ICML 2026, NVIDIA Research revealed a method to make video diffusion models far faster.