AV1 Architecture Basics

AV1 was designed by a consortium with an explicit goal of open usage terms, but its technical story stands on its own. It inherits general ideas from its predecessors and pushes partitioning, prediction and entropy coding further in service of streaming efficiency.

Superblocks and deep partitioning

The base unit is the superblock, up to one hundred and twenty-eight samples square. Within it, AV1 supports an unusually rich set of partition shapes, including rectangular splits that map cleanly onto the wide horizontal or vertical structure common in real footage. Deeper partition trees let the encoder spend its complexity where detail actually exists.

Streaming surfaces motivated many of the design trade-offs

Prediction and filter tools

AV1 adds more directional intra modes and a set of smooth and recursive variants that model gradients efficiently at low rate. Inter prediction gains warped motion, which can describe rotation and shear with a single parametric model instead of many block vectors, and a reference-frame framework that lets the encoder maintain a flexible pool of candidate frames.

After reconstruction, loop filters are applied. AV1 uses a deblocking filter alongside a constrained directional enhancement filter and a restoration stage, each aimed at a different artefact family. The combination improves both objective scores and perceived cleanliness at low rate.

Entropy coding

A multi-symbol arithmetic coder replaces the binary-only approach of earlier designs, and the probability model adapts as coding proceeds. Because symbol alphabets are larger, fewer coding decisions are needed for the same information.

Feature	What it contributes	Practical effect
Superblock partition	Flexible geometry	Efficiency on real-world detail
Warped motion	Parametric global motion	Cheaper for camera movement
Multi-symbol coder	Richer entropy model	Lower overhead per decision
Loop restoration	Artefact suppression	Cleaner gradients at low rate

AV1's encoder complexity is the main practical burden. Processor-based encoding is slow; hardware paths exist but their toolset is a subset of the specification.

Read the ladder planner next to see where AV1 fits in a mixed device population.