How AV1 Preserves Edge Coherence in Scalable Video

Scalable video coding in AV1 allows a single bitstream to provide multiple spatial resolutions, enabling seamless adaptation to fluctuating network bandwidth and diverse display capabilities. Preserving edge coherence across these scalable spatial enhancement layers is critical to preventing visual artifacts such as ringing, blurring, and edge misalignments. AV1 accomplishes this through a combination of normative inter-layer upscaling filters, direction-aware in-loop filtering, fine-grained directional prediction, and specialized block-partitioning mechanisms.

Normative Inter-Layer Upscaling and Phase Alignment

In AV1's scalable framework, an enhancement layer often uses decoded frames from lower-resolution base layers as reference frames. To utilize a lower-resolution picture as a reference, the frame must be scaled up to match the dimensions of the enhancement layer.

AV1 specifies normative polyphase interpolation filters to perform this upscaling. Because the standard strictly defines these 8-tap filter kernels and their fractional phase handling, both the encoder and decoder produce mathematically identical upscaled references. This precise phase alignment prevents spatial phase shifting—a common cause of edge flickering, double edges, and jaggy artifacts when reconstructing high-frequency boundaries from lower-resolution data.

Constrained Directional Enhancement Filter (CDEF)

The Constrained Directional Enhancement Filter (CDEF) plays a central role in maintaining sharp, coherent edges. CDEF operates after the standard deblocking filter and before the loop restoration stage.

  1. Direction Search: CDEF divides blocks into 8x8 regions and identifies the dominant direction of edges within each block from eight possible angles.
  2. Directional Filtering: Once the primary edge direction is determined, a non-linear 1D filter is applied along that direction rather than across it.
  3. Artifact Suppression: By filtering parallel to edges, CDEF removes ringing artifacts caused by coarse quantization in base-layer references without smoothing or dulling the sharp transitions across the edge boundary.

When enhancement layers add residual high frequencies to an upscaled base layer, CDEF ensures that residual noise does not degrade the primary directional contours established across layers.

Directional Intra Prediction and Angle Deltas

When enhancement layers rely on spatial prediction, AV1 offers 56 directional intra-prediction angles. Instead of limiting prediction to standard horizontal, vertical, or broad diagonal angles, AV1 allows an encoder to signal fine-grained angle offsets (deltas).

This precision allows enhancement layers to track continuous diagonal edges across block boundaries. When predicting structural details on top of an upscaled reference, these fine angles prevent stepping artifacts (aliasing) and preserve consistent edge slopes throughout the picture.

Wedge Code and Compound Prediction

AV1 introduces wedge-based prediction, a specialized mode within compound prediction. Wedge prediction splits a coding block into two regions along an arbitrary diagonal or vertical/horizontal edge using a predefined codebook of 16 wedge masks.

In scalable enhancement layers, wedge prediction allows the codec to blend an inter-layer upscaled reference with a temporal reference or an intra-predicted block along an exact edge contour. The smooth transition mask applied at the wedge boundary prevents harsh discontinuities while keeping edge transitions sharp and localized.

Dual Loop Restoration Filters

Following CDEF, AV1 applies Loop Restoration, which provides two switchable filtering algorithms: a separable symmetric Wiener filter and a self-guided filter.

These restoration filters reconstruct the high-frequency spectral components that are inherently attenuated in lower spatial layers, ensuring that edge transitions remain crisp, stable, and visually coherent across layer switches.