From H.264 to High Efficiency Video Coding, AV1 and Versatile Video Coding
H.264, High Efficiency Video Coding (HEVC), AV1 and Versatile Video Coding (VVC) are four formats that share a basic design: predict a block, transform and quantize the residual, entropy code the result, filter the reconstruction. This page compares selected block sizes, prediction choices and filters within that shared design, using the limits one encoder for each format defines in its own source.
Compare the building blocks
Choose a format. The three drawings show the same 128×128-pixel area: how the format's largest unit divides it, the largest square transform the format allows, and a count of the directions intra prediction can take. The largest unit is the macroblock, coding tree unit or superblock that organizes a region; the drawing is not a partitioning an encoder chose. The frame size changes only the counts of units, which are grid counts.
| Compared | In the format, as defines it |
|---|---|
| Largest block | |
| Largest transform | |
| Selected luma intra modes | |
| Entropy coder | |
| In-loop filters | |
| Stream unit |
For a 1920×1080 frame:
- The largest unit is 16 pixels a side in H.264, 64 in HEVC and 128 in AV1 and VVC. The frame is a grid of 8,160 macroblocks, 510 coding tree units, or 135 of the largest units in AV1 or VVC.
- The largest transform is 8×8 in H.264, 32×32 in HEVC and 64×64 in AV1 and VVC: 16 and 64 times the area of 8×8.
- The directional luma intra modes counted are 8, 33, up to 56 and 65. AV1 reaches 56 as 8 named directions, each at up to 7 angles; libaom allows the extra angles on every block size but 4×4, 4×8 and 8×4. Other intra tools are not counted: libaom also lists five filter-intra modes, and VVenC a further set of up to 32 (
MAX_NUM_MIP_MODE). - The newer formats shown use more in-loop filters alongside deblocking: one in HEVC, two in AV1 and two in VVC, and not the same two.
What stayed the same
The stages of the hybrid loop remain recognizable in all four: prediction, residual coding, reconstruction filtering. Their rules differ. H.264 organizes a frame into fixed 16×16 macroblocks with a set of allowed partitions; the newer formats use larger units and their own partition structures, as Block partitioning shows for H.264 and HEVC. Knowing the shared loop helps you find where a tool sits. It does not make the tool's exact operation the same from one format to the next. The table compares coding tools and, in its last row, stream packaging, the subject of Bitstream structure.
What more choices cost
- A richer toolset creates more possible decisions. Encoders control the work through eligibility rules, pruning and presets, so these counts alone do not predict encoding speed.
- A large block can pay off where one prediction describes a region well, and that can be a detailed region moving as one piece. Where the prediction has to change inside a region, smaller partitions can help. The largest size is a limit, not a promise about the blocks an encoder will choose.
- The gain is not on this page. How many bits a newer format saves depends on the encoders compared, their settings, the content and the metric; none was measured here.
Sources
Read 2026-11-15, at the commits linked. The angle-step rule, the further intra modes and VVenC's entropy coder were read 2026-11-16.
- x264 at commit c24e06c (GitHub mirror): the nine 4×4 intra modes in
common/predict.h; the 16×16 macroblock, the 4×4 and 8×8 transforms, the two entropy coders and the deblocking filter are covered, with their own sources, on the earlier pages of this course. - x265 at commit 4191822: in
source/common/common.h, 35 intra modes, of which two are planar and DC, the largest coding unit and the largest transform. - libaom at commit 1b7f9ef (GitHub mirror): in
av1/common/enums.h, the 128×128 block, the count of intra modes, and 8 directional modes with angle steps of up to 3 either way and five filter-intra modes; the block sizes that may use the angle steps inav1/common/reconintra.h, read against the order of the block size list; the largest transform inaom_dsp/txfm_common.h; writing a symbol and the limit of 16 values. - VVenC at commit a455d51: in
source/Lib/CommonLib/CommonDef.h, the tree depth for a 128-pixel unit, the largest transform and 67 luma intra modes and the further set of 32; ininclude/vvenc/vvencCfg.h, the switches for deblocking, sample adaptive offset and the adaptive loop filter; insource/Lib/EncoderLib/BinEncoder.cpp, the routine that codes one binary decision against a context's probability by narrowing a range. - Not checked: the four specifications; partition shapes, inter prediction tools, transform types and anything else not in the table; VVenC beyond the constants, switches and coding routine named; which blocks each format's other intra tools apply to; any measurement of compression gain or of encoding time.