AV1 Under the Hood

AV1 rebuilds the encode path with larger partitions, richer motion and a fresh entropy stage. It reaches the lowest size of the three at a fixed distortion, and it is the most demanding to encode. This sheet sets out the pieces.

Superblocks and the split tree

AV1 starts from a superblock as large as one hundred and twenty-eight samples and splits it into a tree of prediction and transform blocks. The larger top size helps flat regions; the deep splits help fine texture. The encoder searches far more shapes than HEVC, and that search is the main reason AV1 encodes slowly on a general processor.

Figure 1. The AV1 path shares the broad stages of any encoder but widens the search at each one.

Motion and restoration

AV1 adds a richer motion model, including warped motion for zoom and rotation, and an interpolation choice that refines sub-pixel matches. It also adds loop restoration filters that clean artefacts inside the decode loop, so a later frame predicts from a cleaner reference. Each tool shrinks the residual, and a smaller residual survives harder quantisation.

Table 1. AV1 tools and their role.
Tool	Role
superblock	Large top partition
warped motion	Zoom and rotation
loop restoration	Cleaner reference frames
multi-symbol entropy	Packs symbols tightly

Entropy and the decode side

The entropy stage uses a multi-symbol model that adapts to local statistics, which packs bits tightly at the expense of a busy decoder. On mainstream hosts the decode side is now well supported, though the base is more uneven than H.264 or HEVC, so a ladder that leans on AV1 should keep a lower rung in a codec the whole audience can read.

Where it pays

AV1 earns its keep at 1440p and above, where the size saving across a large audience is worth the encode time and the uneven decoder base. Anchors for the top tier are on the 4K settings sheet, and the measurement page covers how the saving is scored.

Comparing AV1 to HEVC on one clip

On a high-resolution clip, AV1 tends to reach a given score at the smallest size of the three, which is why it anchors the top rung. The gain is real but uneven: easy content shows a small gap, while hard content shows a larger one. The desk reports the gap with the clip class, never as a single figure.