AV1 Directional Intra Prediction Modes Explained
This article provides an overview of intra prediction in the AV1 video codec, focusing specifically on its directional modes. AV1 significantly expands upon older video formats by offering a total of 56 directional intra prediction modes. By utilizing a base set of nominal directions combined with fine-grained angular offsets, the codec achieves exceptional spatial compression efficiency while maintaining computational feasibility.
Nominal Directional Modes
At its core, AV1 defines 8 nominal directional modes. These modes represent standard prediction angles designed to capture edge gradients across neighboring decoded pixels:
- V_PRED: Vertical prediction (90 degrees)
- H_PRED: Horizontal prediction (180 degrees)
- D45_PRED: Diagonal prediction at 45 degrees
- D135_PRED: Diagonal prediction at 135 degrees
- D113_PRED: Diagonal prediction at 113 degrees
- D157_PRED: Diagonal prediction at 157 degrees
- D203_PRED: Diagonal prediction at 203 degrees
- D67_PRED: Diagonal prediction at 67 degrees
These nominal modes serve as anchors for directional spatial modeling across blocks.
Angle Deltas and the 56 Modes
To capture arbitrary edge angles without drastically increasing syntax overhead, AV1 introduces angle deltas to its nominal modes. For each of the 8 nominal directional modes, the encoder can signal an angular offset step ranging from -3 to +3 (excluding 0, which corresponds to the base nominal mode). Each step represents an offset of roughly 3 degrees.
With 7 possible variations (the nominal base angle plus 6 offset angles) applied across each of the 8 nominal directional modes:
\[\text{8 nominal modes} \times \text{7 angle steps} = \text{56 directional modes}\]
This provides a continuous sweep of prediction angles from 45 degrees to 207 degrees.
Significance in Compression
Compared to its predecessor VP9, which only supported the 8 base directional modes, AV1's 56 directional modes significantly reduce prediction error (residual energy) for textures with diagonal orientations. This fine angular granularity allows AV1 to achieve higher visual quality at lower bitrates, bridging the gap between complexity and compression efficiency.