Video AI models generate clips with 99% fewer attention computations

At 1% attention density, the method matches or surpasses the quality of previous methods using 12.5% density, while achieving up to 5.39x speedups.

Research Lab

Enhuai Liu · Yunke Wang · Yutong Wang · Changming Sun

The University of Sydney · CSIRO

Research Digest·10 November 2026·2 min read

Thread:Diffusion Inference Acceleration

Illustration: Video AI models generate clips with 99% fewer attention computations

In one line

ESTRA accelerates video diffusion models up to 5.39× by computing only 1% of attention exactly.

Before you read this

Three ideas this paper builds on

sparse attentionSelectively attending to only the most relevant positions or tokens to reduce computational overhead.50 papers →diffusion transformerA neural network architecture combining transformer layers with diffusion processes for high-quality image and video generation.70 papers →kv cache compressionReducing the size of the key-value cache in transformers via selective retention, low-rank updates, or adaptive compression.78 papers →

Read it as

Generating video with AI is slow because the model must compare every part of every frame to every other part. This quadratic scaling makes longer or higher-resolution videos particularly expensive. Existing methods skip most of these comparisons to speed things up, but quality suffers badly when skipping more than about 90% of them.

ESTRA is a new technique that keeps video quality high even when skipping 99% of the comparisons. It groups similar pieces of the video together, computes the important comparisons exactly, and then approximates the rest using group averages. A clever correction improves the approximation's accuracy.

On two large video models, ESTRA generated videos 3.26 to 5.39 times faster than the standard method while preserving quality. At its most aggressive setting, it was still more faithful to the original video than a competing method that used twelve times more computation. This makes high-quality video generation much more practical.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

sparse attention diffusion transformer training-free acceleration variance-calibrated mean compensation token-level selection kv cache compression