Modern Video Codecs

The video codec that wins is almost never the one that compresses best. It is the one that clears two bars the spec sheet does not mention: a patent-licensing regime cheap and predictable enough that device makers will ship it, and a hardware decoder in enough silicon that playback does not melt the battery. HEVC is roughly half the bitrate of H.264 at the same quality and has spent a decade as the format the open web refuses to touch, because its licensing fragmented into competing royalty pools no one could safely pay. AV1 is not dramatically better than HEVC on pure efficiency, yet Google, Netflix, Meta, and Amazon bet the future of streaming on it for one reason: it is royalty-free. The honest study of modern video codecs is therefore not a compression leaderboard. It is the interaction of three forces — rate-distortion efficiency, hardware support, and licensing — where the second and third routinely overrule the first.

Video makes this harder than still images by an order of magnitude, because video has a time axis. A still-image codec exploits redundancy within a frame; a video codec also exploits redundancy between frames, predicting each picture from its neighbors with motion compensation, and that single addition is where almost all the compression and almost all the encoder cost live. Understanding modern video codecs means understanding the temporal game first, then layering on the brutal economics of who is allowed to use which tool. The still-image side of this same format war — AVIF, JPEG XL, WebP — is covered separately in image codecs compared; here the subject is moving pictures, where the stakes and the patents are larger.

Why video is a different problem than images

A raw 4K60 video stream is about 12 gigabits per second. Delivering that over a 25 Mbps connection demands a compression ratio near 500:1, and no intra-frame coding gets close. The leverage comes from the fact that consecutive frames are nearly identical: the background holds still while a few objects move. A video encoder exploits this by coding most frames as differences from other frames.

The unit of structure is the Group of Pictures (GOP), built from three frame types:

• I-frames (intra) are self-contained still images, the only frames decodable alone. They are the seek points and the recovery points, and they are large.
• P-frames (predicted) store only what changed from a previous frame, as motion vectors plus a residual.
• B-frames (bi-directional) predict from both past and future frames, the most efficient and the most computationally expensive.

GOP structure (display order), GOP size 8:

  I  B  B  P  B  B  P  B  B  I ...
  |        ^        ^        |
  |        |        |        next keyframe (seek point)
  keyframe P-frames reference back; B-frames reference both directions

  encode order differs from display order because B-frames
  need their future reference decoded first.

The expensive part is motion estimation: for each block the encoder searches neighboring frames for the best match, producing a motion vector. Newer codecs partition frames into ever-more-flexible block sizes (HEVC’s coding tree units up to 64×64, AV1’s superblocks up to 128×128 with deep recursive splits), offer dozens of intra-prediction directions, and add tools like AV1’s film-grain synthesis (strip the grain before encoding, re-synthesize it on playback) and screen-content modes. Every one of these tools improves compression and costs encoder time, and the search space grows so fast that a high-quality AV1 encode can run hundreds of times slower than real time. That asymmetry — cheap to decode, brutally expensive to encode well — shapes everything downstream.

The four contenders

H.264/AVC is the format everything can decode. Two decades of ubiquity, hardware decoders in every screen ever made, and the lowest encode cost make it the safe default and the lingua franca of video calls and compatibility fallbacks. HEVC/H.265 brought 10-bit HDR, 4K and 8K, and roughly half the bitrate — and walked straight into a licensing swamp. AV1 was the open-source industry’s deliberate answer to that swamp, engineered by a consortium specifically to be royalty-free. VVC/H.266 is the newest and most efficient on paper, the natural HEVC successor, and so far the least adopted, for reasons that are again about money and silicon rather than math.

Measuring “better”: rate-distortion and VMAF

“50% smaller” is a claim that means nothing without specifying quality, content, and encoder. Codecs are compared on rate-distortion (RD) curves — quality plotted against bitrate — and condensed into the BD-rate (Bjøntegaard Delta rate), the average percentage of bitrate saved at equal quality across the curve. A codec with a BD-rate of -30% versus a reference reaches the same quality at 30% less bitrate.

The quality axis itself is contested. PSNR, the old pixel-error metric, correlates poorly with perception and flatters codecs that smooth detail. The industry moved to VMAF (Video Multi-method Assessment Fusion), Netflix’s open-source perceptual metric trained on human scores, now the de facto standard for streaming. You can compute it directly in ffmpeg:

1
2
3

# Compare an encode against the source with VMAF (0-100, higher is better)
ffmpeg -i encoded.mp4 -i source.mov -lavfi libvmaf -f null -
# typical good streaming target: VMAF 93-95 at the top of the ladder

Representative BD-rate savings versus H.264, on typical content at streaming bitrates — directional, because real numbers swing with content and encoder maturity:

Two honest caveats dominate. First, encoder maturity matters as much as the standard: a well-tuned x265 can beat a naive AV1 encode, and AV1’s leap forward came largely from SVT-AV1 maturing, not from the bitstream changing. Second, the gains are content- and bitrate-dependent — high-motion sports, film grain, and screen content each reshuffle the ranking, and at very high bitrates the codecs converge because there is little left to save.

Hardware versus software encoding

Every codec has two encoders that behave like different products. Software encoders — x264, x265, libaom, SVT-AV1 — run on the CPU, expose the full toolset, and reach the efficiency the standard promises, at the cost of speed. Hardware encoders — NVIDIA NVENC, Intel Quick Sync and Arc, AMD AMF, Apple VideoToolbox — are fixed-function ASIC blocks that encode in real time at tiny power, and trade efficiency for that speed: at the same bitrate a hardware encode generally looks worse than a slow software encode, often by the equivalent of a half-generation of codec progress.

1
2
3
4
5
6
7
8
9

# Software AV1, quality-targeted (slow, most efficient)
ffmpeg -i in.mov -c:v libsvtav1 -crf 30 -preset 6 -c:a libopus out.mp4

# Software HEVC and H.264, CRF mode
ffmpeg -i in.mov -c:v libx265 -crf 24 -preset slow out_hevc.mp4
ffmpeg -i in.mov -c:v libx264 -crf 23 -preset slow out_h264.mp4

# Hardware AV1 on a recent NVIDIA GPU (real-time, less efficient per bit)
ffmpeg -i in.mov -c:v av1_nvenc -cq 30 -preset p6 out_hw.mp4

The rule of thumb: use software at a slow preset for content you encode once and serve millions of times (a streaming catalog, a video-on-demand title), where CPU hours amortize into bandwidth saved; use hardware for live, real-time, and high-volume user-generated uploads where latency and throughput dominate. The CRF number is the main quality dial in software (lower is higher quality and bigger files); presets trade encode speed against efficiency.

Crucially, decode is the gating factor for adoption, not encode. A codec can have a great software encoder and still fail in the market if phones, TVs, and browsers lack a hardware decoder, because software decoding 4K AV1 drains a battery and stutters on weak SoCs. AV1’s tipping point was not its encoder; it was hardware AV1 decode arriving in mainstream phone and TV chips around 2020-2022. VVC’s problem is precisely the absence of that silicon.

The patent landscape that decides adoption

This is the axis that has decided every recent outcome, and it is worth being specific.

H.264 is licensed through a single pool (MPEG LA, now Via LA) with a famous, decisive concession: video distributed free to end users over the internet carries no royalty. That clarity is a large part of why H.264 became universal — a site could stream it without a licensing department.

HEVC shattered that model. Its patents are spread across multiple competing pools — Via LA, Access Advance (formerly HEVC Advance), Velos Media — plus licensors who joined no pool at all, and some pools sought royalties not just on devices but on content and streaming. No one could state with confidence what it cost to ship and stream HEVC, so browser vendors and web platforms declined, and HEVC was confined to the walled gardens that could negotiate it: Apple’s ecosystem, broadcast, and UHD Blu-ray. The lesson was unmistakable and it created AV1.

AV1 was built by the Alliance for Open Media — Google, Netflix, Amazon, Meta, Microsoft, Apple, Intel, Mozilla, Nvidia and others — explicitly to be royalty-free, with a defensive patent cross-license among members. That is why the companies that actually move the most video on Earth could deploy it without negotiation, and why it now carries a large and growing share of YouTube, Netflix, and Meta delivery. Its efficiency edge over HEVC is real but modest; its licensing edge is the whole game.

VVC/H.266 is repeating HEVC’s mistake. It is the most efficient option on paper, but its patents are pooling under Access Advance and Via LA with no royalty-free path, and hardware support is scarce. The result is a superior codec stuck in the same trap that strangled its predecessor — strong adoption in some broadcast and niche professional contexts, near-invisibility on the consumer web. The exact same dynamic played out in audio codecs, where the royalty-free option won distribution despite not always winning the benchmark.

Where each actually wins

The decision collapses cleanly once licensing and hardware support are weighed alongside efficiency.

The nuances: H.264 wins anywhere compatibility or latency beats efficiency — a video call cannot wait for a slow encoder, and a fallback stream must play on a decade-old TV. HEVC wins inside ecosystems that already licensed it, especially Apple devices, broadcast 4K, and physical UHD media, and it remains the practical HDR workhorse where AV1 hardware is absent. AV1 wins the open, high-volume streaming case decisively, which is why the largest video platforms drove it. VVC wins only the narrow professional and broadcast slices willing to pay its royalties, unless and until its licensing clarifies. And whichever codec you choose, the HDR and wide-gamut signaling has to be carried correctly end to end or “more color” becomes “wrong color,” the failure mode detailed in color management across cameras, screens, and prints.

A practical encoding playbook

For adaptive streaming you do not pick one bitrate; you build an encoding ladder of multiple resolution/bitrate rungs and let the player switch based on bandwidth. The modern refinement is per-title (or per-shot) encoding, popularized by Netflix: analyze each title’s complexity and tailor the ladder, because a flat cartoon and a grainy action film need wildly different bitrates for the same VMAF. This is where the bandwidth savings actually land in production, and it interacts directly with delivery economics — the cache behavior and origin offload of a modern CDN and the broader bandwidth-versus-quality math of web performance engineering.

For low-latency live, the levers reverse: shrink the GOP, drop or limit B-frames, and prefer hardware encode to keep latency bounded, accepting the efficiency hit. A representative low-latency H.264 command:

1
2
3

# Low-latency live H.264: small GOP, no B-frames, fast tune
ffmpeg -i input -c:v libx264 -preset veryfast -tune zerolatency \
  -g 60 -bf 0 -b:v 4500k -maxrate 4500k -bufsize 9000k out.ts

In professional capture and broadcast contexts the codec choice also has to survive multiple encode generations and live switching, the workflow constraints discussed in live broadcast cameras and switchers, where intra-only or lightly-compressed mezzanine codecs often beat the most efficient delivery codec because they cut and re-encode cleanly.

Verdict

Modern video codecs sort not by who compresses best but by who is allowed to ship and who has silicon to decode. H.264 remains the universal baseline and the right answer wherever compatibility or low latency matters more than bitrate — it plays everywhere and encodes cheaply, and that is worth more than efficiency in a video call or a fallback rung. HEVC is technically excellent and commercially hobbled, thriving only inside the walled gardens — Apple, broadcast, UHD Blu-ray — that could absorb its fractured licensing, and serving as the practical HDR workhorse where AV1 hardware has not yet arrived. AV1 is the open web’s winner, not because its efficiency lead over HEVC is large but because it is royalty-free and now hardware-decoded at scale, which is exactly why the companies that move the most video built it. VVC is the cautionary sequel: the best numbers on the page, trapped in the same licensing-and-silicon swamp that confined HEVC, and absent a royalty-free path it will stay a broadcast-and-niche format no matter how good the compression is.

The durable lesson is the one the patent column keeps teaching: efficiency proposes, licensing and hardware dispose. Measure with VMAF rather than PSNR, encode once in slow software for catalogs and in hardware for live and high-volume uploads, build per-title ladders rather than flat bitrates, and choose the codec by where it will actually be decoded and whether you can legally afford to stream it. Do that and the codec question stops being a benchmark argument and becomes what it really is: a deployment decision constrained by silicon and law.

Sources

• ITU-T H.264, “Advanced video coding for generic audiovisual services” — https://www.itu.int/rec/T-REC-H.264
• ITU-T H.265, “High efficiency video coding (HEVC)” — https://www.itu.int/rec/T-REC-H.265
• ITU-T H.266, “Versatile video coding (VVC)” — https://www.itu.int/rec/T-REC-H.266
• Alliance for Open Media, “AV1 Bitstream & Decoding Process Specification” — https://aomedia.org/av1/specification/
• Netflix Technology Blog, “VMAF: The Journey Continues” — https://netflixtechblog.com/vmaf-the-journey-continues-44b51ee9ed12
• Netflix Technology Blog, “Per-Title Encode Optimization” — https://netflixtechblog.com/per-title-encode-optimization-7e99442b62a2
• SVT-AV1 encoder project — https://gitlab.com/AOMediaCodec/SVT-AV1
• Access Advance, “VVC Advance Patent Pool” — https://accessadvance.com/licensing-programs/vvc-advance/
• FFmpeg, “Encoding for streaming and H.264/HEVC/AV1 guides” — https://trac.ffmpeg.org/wiki/Encode/AV1