nablaNABLA: Neighborhood Adaptive Block-Level Attention
Dmitrii Mikhailov, Aleksey Letunovskiy, Maria Kovaleva, Vladimir Arkhipkin, Vladimir Korviakov, Vladimir Polovnikov, Viacheslav Vasilev, Evelina Sidorova, Denis Dimitrov
2025-07-25
Summary
This paper talks about NABLA, a new way to make video transformers faster by using a smart attention method that focuses on blocks of the video and adapts to the important parts dynamically.
What's the problem?
Video generation needs a lot of computing power because paying attention to every small detail in every frame is very slow and expensive, especially for long and high-quality videos.
What's the solution?
The researchers designed NABLA to look at groups of pixels or tokens in blocks and decide which blocks need more focus, adapting the attention dynamically based on what’s important in the video. This reduces how much computing is needed while keeping the video quality very high.
Why it matters?
This matters because NABLA speeds up video generation by almost three times without making the videos look worse, making it easier and cheaper to create high-quality videos with AI.
Abstract
NABLA, a Neighborhood Adaptive Block-Level Attention mechanism, enhances video diffusion transformers by reducing computational overhead without significantly impacting generative quality or visual fidelity.