Breaking

NVIDIA's FAST-AR Video Diffusion Speedup Accepted at ICML 2026

July 7, 2026 at 10:12 EDT

Research & Papers
Media Generation
Foundation Models

NVIDIA Research's Dvir Samuel and colleagues have unveiled FAST-AR, a training-free framework that speeds up autoregressive video diffusion models by up to 10x, and it has been accepted at ICML 2026. The method lightens the attention computation that bottlenecks long-form video and world-model generation.

ICML 2026 · NVIDIA Research

FAST-AR Makes Autoregressive Video Diffusion Up to 10× Faster

A training-free, plug-and-play inference framework that compresses the KV cache and sparsifies attention — cutting latency while keeping GPU memory nearly constant across long video rollouts.

5–10×

End-to-end speedup vs dense attention

2,880

Frames (2 min @ 24fps) at near-constant memory

Retraining required — plug-and-play

End-to-end speedup, drawn to scale

Relative processing speed — taller is faster

1×

Dense
attention

up to 10×

FAST-AR

Three modules resolve the attention bottleneck

TempCache

Detects near-duplicate keys to compress & merge the KV cache

→

AnnCA

ANN matching narrows cross-attention to frame-relevant tokens

→

AnnSA

ANN buckets restrict self-attention to semantic neighborhoods

Long-form generation: the payoff

Dense attention

KV cache grows without bound → rising latency, growing GPU memory, degrading temporal consistency.

FAST-AR

Peak GPU memory stays nearly constant even across a 2-minute, 2,880-frame rollout — with quality roughly maintained.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.