How AV1 Signals Skipped Transform Blocks

The AV1 video codec conserves data by identifying areas where prediction is accurate enough that transmitting residual error is unnecessary. To prevent wasting bits on empty transform coefficients, AV1 employs a hierarchical signaling mechanism that flags skipped transform blocks at both the block and sub-transform levels. By utilizing syntax elements like skip_mode, the skip_txfm flag, and context-adaptive entropy coding, the bitstream completely bypasses the transmission of coefficient tokens and transform operations for these regions.

Block-Level Prediction Skip (skip_mode)

At the coding block level, AV1 can activate skip_mode for inter-predicted blocks under specific conditions—namely, when a block uses compound prediction from the two nearest reference frames spanning past and future temporal directions. When skip_mode is set to 1 in the bitstream, the decoder assumes motion vectors derived directly from neighboring blocks and infers that there is no residual data. In this scenario, all transform information is omitted from the bitstream, requiring only a single syntax element to represent the entire block.

Transform Residual Skip Flag (skip_txfm)

When a block does not qualify for or use skip_mode, it can still omit transform residuals if the quantized coefficients are all zero. The bitstream conveys this through the skip_txfm syntax element (equivalent to a coded block flag).

If skip_txfm is true (value of 1):

• The bitstream terminates residual parsing for the block immediately.
• The transform tree syntax, transform type selection (e.g., DCT, ADST, identity), and coefficient tokens are entirely skipped.
• The decoder adds the prediction directly to the reconstruction buffer without performing an inverse transform.

If skip_txfm is false (value of 0), the decoder parses the transform partition tree and extracts transform units to reconstruct the residual signal.

End of Block (eob) Signaling Within Sub-Blocks

When a transform block is parsed, AV1 divides it into smaller transform units (TxUnits) according to the transform tree. If an individual transform unit within an active block contains only zeros or sparse coefficients, AV1 utilizes an end-of-block (eob) position marker.

The eob value signals the index of the highest-frequency non-zero coefficient in the scanning order. If a transform unit has no coefficients at all, eob is signaled as 0, immediately skipping the coefficient payload for that unit. If only low-frequency coefficients are present, signaling an early eob allows the bitstream to omit all subsequent trailing zero coefficients without coding individual zero flags.

Context Modeling for Maximum Bit Conservation

The signaling of skip flags is optimized through AV1's context-adaptive multi-symbol arithmetic coder (based on the Daala entropy coder). The probability context used to code both skip_mode and skip_txfm depends on the skip status of neighboring blocks immediately above and to the left:

• If both neighbors skipped their transform coefficients, the current block has a very high statistical probability of skipping as well.
• The arithmetic coder uses this skewed probability distribution to encode the skip flag using a fraction of a bit.

Through this combination of early syntax exits and neighbor-dependent entropy coding, AV1 ensures that non-informative, all-zero transform blocks consume the absolute minimum bitstream budget.