H.264 / AVC Video

H.264, also called AVC, defines a video coding format used for compressed moving pictures.

H.264, AVC, and MPEG-4 Part 10 name the same video coding standard. It defines the instructions a decoder follows to reconstruct compressed pictures, rather than a particular encoder's recipe for choosing those instructions.

A film in an MP4 file, a camera clip in another container, and a television stream can all carry AVC. Their packaging supplies the coded track to the same kind of reconstruction process.

An Emby session can show AVC at both input and output when the server has decoded and encoded a new AVC representation. The reconstruction steps below explain what differs from copying the encoded track.

The coded data comes in units

AVC's Network Abstraction Layer units, or NAL units, organize the bitstream. Some carry coded picture regions called slices. Others supply configuration or additional information.

A coded picture can contain several slice units. The media packet or container sample that carries them is a packaging boundary; the AVC units describe the coding boundary.

Before reconstructing a slice, the decoder needs the applicable sequence parameter set (SPS) and picture parameter set (PPS). These describe such things as the coded grid, permitted sample format, picture-order behavior, and relevant coding options. A slice identifies the parameters and coding decisions it uses.

This is why extracting one piece of picture data without its configuration does not necessarily provide a complete decoder input.

Divide the picture into reconstruction work

AVC's basic macroblock footprint covers 16 × 16 brightness positions. The picture is divided into those regions, with further partitions for prediction and transforms.

For a progressive 1920 × 1080 presentation, the coded macroblock grid is 120 columns by 68 rows: 8,160 macroblocks, representing a 1920 × 1088 grid before display cropping. The crop describes the visible picture boundary; the coding grid determines this part of the decoder's work.

Different regions can use different prediction choices. An edge, a flat area, and a moving object need not receive the same description.

Construct a predictor

An intra-coded region predicts from available reconstructed samples in the same picture. For example, values along an upper boundary can be extended in a chosen direction to form a candidate block.

An inter-coded region identifies a decoded reference picture and motion information telling the decoder where to sample it. Fractional positions use the interpolation rules specified by AVC. A region can combine permitted reference information according to its coding type and mode.

The encoder searches for a useful predictor. The decoder instead follows the predictor encoded in the syntax. Its task is deterministic reconstruction from the supplied information.

Reconstruct the correction

The source region usually differs from its predictor. The encoder represents that difference as a residual and can transform it into coefficients describing patterns within blocks.

The decoder entropy-decodes those coefficients, reconstructs their scale from the quantization parameters, and applies the specified inverse transform. It then adds the resulting correction to the predictor.

AVC uses integer transform operations for common 4×4 residual blocks; selected profiles permit additional tools such as an 8×8 transform. Quantization can discard precision, so the decoded correction need not equal the original source difference. The compression walkthrough demonstrates that approximation numerically.

The entropy-coding mode is another syntax choice. CAVLC and CABAC use different coding procedures, so the decoder needs the selected mode even when both ultimately reconstruct the same kind of sample array.

Filter and retain the reconstructed picture

AVC can apply its in-loop deblocking filter to reduce selected discontinuities between coded regions. “In-loop” matters: the appropriately filtered reconstruction can become a reference for later pictures.

The encoder reconstructs the same result locally when preparing those later predictions. Otherwise it would predict from samples the decoder does not have.

The decoded-picture buffer retains references and pictures waiting for their display times. An anchor can be decoded before a B-picture that will appear earlier on screen. Picture-order information and presentation timing restore the intended viewing sequence after the dependency work.

AVC profiles, x264, and container labels

Profiles restrict this tool/sample set. Baseline, Main, High, and High 10 are different permitted AVC representations. Levels constrain picture size, throughput, buffering, and other resources.

Label	Place in the process
H.264 / AVC	The bitstream and reconstruction standard
x264	One encoder implementation choosing valid AVC instructions
High / High 10	Profile constraints on tools and sample formats
A container's AVC track identifier	The wrapper's description of its coded video

An x264-made track can be decoded by another conforming AVC implementation. The encoder's search effort and quality settings change which valid representation it creates, while the decoder follows that representation.