Delta-QP Signaling per Superblock in AV1

This article provides an overview of delta quantization parameters (delta-QP) in the AV1 video codec and examines how these values are signaled at the superblock level. It covers the role of delta-QP in adaptive quantization, the frame-level configuration required to enable the feature, and the syntax mechanisms used to transmit local quantization adjustments to individual superblocks.

What is Delta-QP?

In digital video compression, the Quantization Parameter (QP) controls the trade-off between visual fidelity and bitrate. While a video frame is assigned a base QP (base_q_idx in AV1), visual content within that frame often varies significantly in complexity and perceptual sensitivity.

Delta-QP refers to a local offset added to or subtracted from the baseline frame QP. By varying quantization across different regions, an encoder can apply finer quantization (lower QP) to perceptually sensitive areas, such as human faces or detailed textures, and coarser quantization (higher QP) to complex motion or background regions. This local adaptation is critical for psycho-visual optimizations, Region of Interest (ROI) coding, and fine-grained rate control.

Frame-Level Prerequisites in AV1

Before delta-QP values can be signaled for individual superblocks, the encoder must configure the delta-QP mechanism within the frame header:

delta_q_present_flag: A 1-bit syntax element in the frame header indicating whether delta-QP signaling is active for the current frame. If set to 0, all superblocks strictly inherit the frame's base_q_idx.
delta_q_res: Specifies the step size or resolution of the delta values. It dictates the bit-shift applied to the decoded delta value, defining the granularity at which QP offsets are adjusted.

Signaling Delta-QP per Superblock

AV1 operates on superblock units of either 128×128 or 64×64 pixels. When delta_q_present_flag is enabled, the encoder can transmit delta-QP syntax elements to modify the quantization index dynamically throughout the frame.

1. Predictive Coding Scheme

AV1 does not transmit an absolute QP index for every superblock. Instead, it uses a predictive mechanism where the signaled delta represents the difference relative to the previous superblock's active QP value. At the start of a tile, the active QP initializes to the frame's base_q_idx. Each signaled delta updates a running state variable (current_q_index), which carries over to subsequent blocks until another delta is signaled.

2. Triggering Conditions

To conserve bits, delta-QP is not transmitted unconditionally in every superblock. A delta-QP value is only signaled at the first transform block within a superblock that contains non-zero transform coefficients (coded residual data). If an entire superblock is skipped or contains only non-coded blocks, the transmission of delta-QP is deferred, and the superblock retains the current running QP value.

3. Syntax Elements

When the conditions to transmit delta-QP are met, AV1 reads two primary syntax elements:

delta_q_abs: Encoded using a combination of unary and fixed-length entropy coding to represent the absolute magnitude of the offset.
delta_q_sign_bit: Transmitted only if delta_q_abs > 0 to indicate whether the offset is positive or negative.

4. QP Reconstruction Formula

Once decoded, the final quantization parameter index for the superblock is computed:

delta_q = delta_q_sign_bit ? -delta_q_abs : delta_q_abs
current_q_index = Clip3(0, 255, current_q_index + (delta_q << delta_q_res))

The resulting current_q_index is clamped to the valid range of 0 to 255. This value serves as the actual QP index used to derive the AC and DC quantization steps for luma components in the current superblock and establishes the reference baseline for subsequent delta updates.