Updated October 2026: a short answer on the three Shannon limits with formulas, AV2 v1.0 (published 2026) and the July 2026 JVET call for proposals beyond VVC, re-checked neural-codec numbers, and a corrected estimate of how much room is left.

Why this matters

Short answer. The Shannon limit is the hard mathematical ceiling Claude Shannon proved in 1948: no scheme can send data over a noisy channel faster than its capacity C, and no codec can compress a source below its entropy H without losing information. For lossy video the limit is the rate-distortion curve R(D), the fewest bits per second needed for a given quality. Nobody can compute that curve exactly for real video, but in 2026 conventional tools still find about 25 percent over VVC, so the floor is not yet in sight.

If you operate a streaming service, an OTT platform, an e-learning library, or a video-conferencing product, your CDN bill scales linearly with bitrate and your viewer experience scales non-linearly with quality. Every percentage point a new codec saves at the same quality is real money. But you cannot plan around vendor marketing: a slide claiming "70 percent better than AV1" needs to be read against the theoretical maximum any codec can ever achieve, otherwise you size your encoding farm for a number that will never arrive. This article gives you the tools to read those claims — what Shannon's limit actually is, what it does and doesn't bound, where conventional and neural codecs sit on the curve in 2026, and how to estimate, for your own content and your own quality target, how much room is left.

What is the Shannon limit? Three limits, one name

"Shannon limit" is used for three related bounds from Shannon's 1948 paper and his 1959 follow-up on rate-distortion. In telecom it means channel capacity; in compression it means entropy (lossless) or the rate-distortion function (lossy). Video codecs live on the third one.

Limit What it bounds Formula Where it matters in video
Channel capacity (Shannon–Hartley) Maximum error-free data rate over a noisy channel C = B · log2(1 + S/N) bits/s Wi-Fi, 5G, cable and satellite links that carry the stream
Entropy (source coding theorem) Minimum bits per symbol for lossless compression H = −Σ p · log2 p Entropy coders (CABAC, arithmetic coding), FFV1 and other lossless archives
Rate-distortion function R(D) Minimum bitrate for a given average distortion Gaussian source: R(D) = ½ log2(σ2/D) bits/sample Every lossy codec: H.264, HEVC, AV1, VVC, AV2, neural codecs

The Gaussian formula gives a rule of thumb codec engineers still use: on an idealised source, every extra bit per sample buys about 6.02 dB of PSNR, and halving the mean squared error (+3 dB) costs half a bit per sample. Real video is far more predictable than a Gaussian, which is why codecs land well below that line, and also why its exact floor cannot be written down.

What Shannon actually said in 1948

Claude Shannon's 1948 paper A Mathematical Theory of Communication did two related things. First, he defined a number called entropy, written H, that measures the average information content of a source — the average number of bits per symbol that the source produces, given how its symbols are distributed.1 The phrase "average information content" is doing real work here. If a source always emits the same symbol, its entropy is zero — the sequence is fully predictable, so describing it costs zero bits per symbol on average. If a source emits one of 256 symbols with equal probability, its entropy is 8 bits — every symbol is maximally surprising.

Second, Shannon proved the source coding theorem: it is impossible to compress a source losslessly below its entropy without losing information, and any compression scheme that uses fewer bits per symbol than H must, on average, fail to reconstruct the source.2 You can get arbitrarily close to H — modern entropy coders like arithmetic coding and CABAC sit within a fraction of a percent of the limit — but you cannot go beneath it. That is the lossless floor.

A useful analogy. A locked filing cabinet has a key, and the key has a minimum number of bits required to describe its shape. You can package the key in a smaller envelope by folding it cleverly, but the key itself is exactly N bits long. You cannot shrink the key — only the envelope.

For video, the lossless floor matters in archival and medical imaging, where formats like FFV1, ProRes 4444, and JPEG-XL operate, but it is not where consumer codecs live. Consumer codecs are lossy, which means they get to throw information away — and that opens a different door.

The lossy door: rate-distortion theory

In the same body of work, Shannon defined a second limit that does apply to consumer video. The rate-distortion function R(D) gives the smallest possible bitrate R at which a source can be encoded such that the average reconstruction error stays below a target distortion D.3 R(D) is a curve, not a single number. As you allow more distortion (D rises), the required bitrate (R) falls. As you demand near-perfect reconstruction (D approaches zero), R climbs toward the entropy H.

The shape of the R(D) curve depends on three things. First, what the source actually looks like — a low-motion talking-head shot has a different curve from a 4K sports broadcast, and a piece of computer-generated animation has yet another. Second, how you measure distortion — mean squared error gives one curve, structural similarity gives another, VMAF gives a third. Third, the statistical model of the source — a richer model, capturing temporal and spatial correlations, produces a tighter (lower) R(D) curve.

R(D) is the right object to think about when you compare codecs. Every codec, conventional or neural, can be plotted as a point or a curve in (R, D) space. The Shannon limit is the lower-left envelope: the curve below which no codec can go for that particular source and that particular distortion measure. Generational progress in codecs is the slow walk of the codec curve down toward the Shannon limit.

A line chart titled Codec generations walk toward the R(D) floor. The horizontal axis is bitrate at the same quality relative to H.264 on a log scale, marked 0.05, 0.1, 0.2, 0.5, 1 and 2; the vertical axis is quality, low to high. Six curves rise from left to right, each with its own colour and dash pattern: a thick solid Shannon limit R(D) on top, then DCVC-FM (2024), VVC (2020), AV1 (2018), HEVC (2013) and H.264 (2003). The empty area above the top curve is labelled No codec can reach this region. A dashed horizontal line labelled Same quality target crosses all six, and a dot marks each crossing: the Shannon limit at 0.09, DCVC-FM at 0.18, VVC at 0.25, AV1 at 0.35, HEVC at 0.50 and H.264 at 1.0, so each generation reaches the same quality further to the left. A note says the curves are schematic, that each codec sits at its measured BD-rate saving against H.264 at the same quality, that DCVC-FM is neural while the other four are conventional standards, and that the gap to the floor shrinks each generation, and so does the room left to gain Figure 1. Codec progress reads as a walk toward the rate-distortion floor. Each generation closes part of the remaining gap; the floor's position is schematic, not a measured bound.

Where conventional codecs sit on the curve in 2026

The most-cited generational milestones in commercial codecs measure relative bitrate savings using the Bjøntegaard delta rate, abbreviated BD-rate. BD-rate gives the average bitrate change between two codecs at the same objective quality, integrated across a range of quality points.4 The numbers below are BD-rate savings against H.264, measured on standard JVET test sets, mostly at HD and UHD resolutions.

HEVC (2013) buys roughly 50 percent over H.264 on standardised test content, with the win larger on UHD than HD because HEVC's larger coding tree units handle high-resolution flat areas more efficiently.5 AV1 (2018) lands at roughly 30 percent over HEVC on the same content, or about 65 percent over H.264.6 VVC (2020) measures about 40 to 50 percent over HEVC on UHD content, which compounds to roughly 75 percent over H.264.7 At 8K resolution, University of Žilina measurements published in Electronics in March 2024 land VVC at 78 percent savings, AV1 at 63 percent, and HEVC at 53 percent — all relative to H.264.8

AV2, whose v1.0 specification AOMedia published in June 2026, adds about 30 percent over AV1 at the same quality; AOMedia's common-test-condition results from November 2025 put the gain at 28.6 percent on YUV-PSNR and 32.6 percent on VMAF.9 That compounds to roughly 75 percent over H.264 — almost identical to H.266/VVC. The next ITU/MPEG step is already in motion: JVET's Enhanced Compression Model (ECM) reached about 28 percent over VVC by its version 18 at the end of 2024, and in July 2026 JVET issued a joint call for proposals for compression beyond VVC, with responses due for evaluation in January 2027 and a standard expected around the end of 2029.17

A useful pattern emerges. H.264 to HEVC was a 50 percent halving. HEVC to AV1 was a 30 percent third. AV1 to AV2 is another 30 percent third. The absolute bitrate saved shrinks each generation, because each 30 percent comes off a smaller base. The relative step per generation stays near 30 percent, but each step now costs far more encoder and decoder complexity than the last.

Codec Year BD-rate vs H.264 Notes
H.264 / AVC 2003 0% (baseline) Workhorse of the internet
HEVC / H.265 2013 -50% Patent thicket slowed adoption
AV1 2018 -65% Royalty-free; about 30% of Netflix streaming (Netflix, Dec 2025)
VVC / H.266 2020 -75% Best conventional standard; weak market traction
AV2 2026 about -75% ~30% over AV1; v1.0 spec published June 2026, software decoder dav2d in progress

The pattern is the diminishing-returns curve of any maturing technology. We are not at the floor — there is still room, especially on high-motion and high-resolution content — but the room is thinner each cycle.

The new dimension: perception

For a long time, the codec community measured progress along the (R, D) axes using objective distortion measures like PSNR or SSIM. The dirty secret of those measures is that they correlate poorly with what a human viewer actually sees. A picture can have low MSE and look terrible (waxy, over-smoothed faces) or high MSE and look great (realistic film grain, sharp textures). Netflix's VMAF was a major step forward because it correlates better with subjective scores, but it is still a per-frame regression onto a fixed observer model.

In 2018 and 2019, Yochai Blau and Tomer Michaeli published a series of papers that re-framed the problem in a way that has become central to the field.10 They showed that distortion and perceptual quality are mathematically at odds — at a fixed bitrate, you cannot simultaneously minimise the mean squared error and the divergence between the distribution of reconstructed images and the distribution of natural images. The two objectives trade off, and the trade-off has a precise rate-distortion-perception (RDP) function.11

The practical consequence: at a given bitrate, the picture can be optimised to look correct ("plausible image from the same distribution") at the cost of being slightly wrong ("each pixel differs from the original more than necessary"). Or it can be optimised to be pixel-accurate, at the cost of looking subtly degraded. The two cannot both be minimised. Modern neural codecs, which can be trained against any differentiable loss, have started to exploit this — they sit on a different part of the RDP surface than conventional codecs do.

A two-dimensional chart titled The rate-distortion-perception trade-off, subtitled that at one fixed bitrate a codec can be pixel-accurate or look natural, not both. The horizontal axis is per-pixel distortion D, low to high; the vertical axis is perceptual divergence P, low to high; a note says bitrate R is held fixed. One convex curve falls from the top left to the bottom right. The shaded area below and left of the curve is labelled Not reachable at this bitrate, and the open area above it is labelled Reachable, but worse on both axes. A blue dot high on the left is labelled Conventional target, HEVC, AV1, VVC, AV2; a purple dot low on the right is labelled Neural target, DCVC-FM, DCVC-RT. Two panels below carry the same point: conventional codecs optimise the per-pixel error, so the picture matches the source but can look waxy, while neural codecs match the distribution of natural video, so it looks right and each pixel differs more Figure 2. The Blau-Michaeli rate-distortion-perception surface. Conventional codecs optimise distortion; neural codecs can shift along the perception axis at the same bitrate.

Where neural codecs stand in 2026

A neural video codec replaces some or all of the hand-designed blocks in a conventional encoder — transform, prediction, entropy coder — with a learned model. The learned model is trained directly on the rate-distortion (or rate-distortion-perception) objective using stochastic gradient descent, which means it can exploit statistical regularities in natural video that hand-coded blocks miss.

Microsoft Research's DCVC family is the most-cited line of work. DCVC-DC (2023) was the first neural video codec to beat VVC's reference implementation VTM on standard JVET test sets under intra-period 32. DCVC-FM (2024), with feature modulation and multi-scale temporal context, beats VTM by about 22 percent in BD-rate under the more demanding intra-period -1 setting (a single intra frame at the start, then unlimited inter prediction), as re-measured in Microsoft's 2025 DCVC-RT paper.12 DCVC-RT (CVPR 2025) is the first practical real-time neural codec: 21.0 percent BD-rate gain over VTM, with 1080p encoding at 125 fps and decoding at 113 fps on an NVIDIA A100 (119 and 105 fps on an RTX 4090).13

Independent benchmarks are more sobering. A University of Bristol low-delay comparison at VCIP 2024 found ECM (the JVET experimental codec that sits above VVC) 16.1 percent ahead of AVM (the AOMedia research codec that became AV2) and 11.0 percent ahead of DCVC-FM on PSNR.14 The same group's April 2026 review reached the same verdict under both low-delay and random-access settings: neural codecs are competitive with, and sometimes better than, VTM and AVM, but still trail ECM.15 On perceptual measures such as LPIPS or FID, neural codecs trained on those losses look better than these PSNR numbers suggest.

The catch in 2026 is the same catch neural codecs have had since they appeared: compute. A neural decoder requires a GPU or a substantial NPU, runs at orders-of-magnitude higher power than a hardware HEVC decoder, and is not yet present in any consumer device's silicon. DCVC-RT's real-time numbers come from an NVIDIA A100 data-centre GPU and an RTX 4090, a 450-watt desktop card. That is fine for VOD pre-encoding (where you spend GPU once and recoup it on every play) and for cloud transcoding, but it is not yet a delivery codec.

How far is the Shannon limit, really?

Short answer: nobody knows the exact number, and anyone who quotes one is using a model, not the source. R(D) can be computed in closed form only for idealised sources such as a Gaussian. For real video, the true statistics are unknown, so every "distance to the Shannon limit" is the distance to a model's bound. What can be measured is the best demonstrated codec: in 2026 that is ECM, which needs about 25 to 28 percent fewer bits than VVC at the same PSNR and stays ahead of the neural codecs in independent benchmarks. So on conventional metrics the floor sits at least that far below today's deployed standards, and probably further.

Two important caveats. First, that floor is not a single number — it depends on the distortion measure. On PSNR, the demonstrated room beyond VVC and AV2 is roughly a quarter of the bitrate. On perceptual measures like VMAF or LPIPS, the floor is lower (you can be further from the original pixel by pixel and still look identical), and there is more room. Second, the RDP surface introduces a third axis — perception — and the floor there is harder to estimate because the "perceptual divergence" objective has many possible definitions, each with its own bound.

The practical implication is that the question "how much more compression is left" splits into three:

  • On PSNR and conventional MSE: about 25 percent more beyond VVC and AV2 is already demonstrated by ECM and targeted by the July 2026 JVET call for proposals, at a steep cost in encoder and decoder complexity.
  • On perceptual measures like VMAF and LPIPS: more room than on PSNR, mostly accessible to neural or hybrid codecs trained on those losses; there is no agreed estimate of how much.
  • On the RDP perception axis: large remaining room, but the question shifts from "how few bits" to "how natural does the reconstruction look", which is a different objective altogether.

A 70 percent compression-improvement headline in 2026 is usually measured on a perceptual axis that conventional codecs were never optimised for. That is real — neural codecs really do produce more natural-looking video at the same bitrate — but it is not comparable with a PSNR-based BD-rate measured under the JVET or AOM common test conditions.

A common mistake: treating "Shannon limit" as a single number

The most expensive misconception in this area is the assumption that there is one Shannon limit you can quote for "video", the way there is one speed of light. There is not. The rate-distortion function R(D) depends on:

  • The source. A 4K sports stream has a different R(D) curve from a low-motion lecture video. Animated content has a curve different from both.
  • The distortion measure. PSNR, SSIM, VMAF, and LPIPS each yield different curves on the same source.
  • The statistical model. A simple block-IID Gaussian model gives one R(D); a temporally-correlated mixture model gives a much lower R(D); a deep autoregressive prior gives lower still.

So the right way to read a vendor claim is: "vs which codec, on which content, with which quality metric, and at which bitrate operating point?" A claim that reads "30 percent better than AV1" with no further context is approximately a coin flip — for some content and some metric it will be true, for some it will not. The Bjøntegaard methodology was invented to standardise this conversation, and the JVET common test conditions and AOM's CTC define exactly which sequences and which configurations are used in academic comparisons.16 If a vendor claim doesn't reference one of these test conditions, the claim is marketing, not measurement.

A worked example: the Shannon-limit budget for a 4K SDR catalogue

To anchor the numbers, imagine a streaming service with a 10,000-hour 4K SDR catalogue, encoded today in HEVC Main 10 at an average 12 Mbps. The catalogue weighs 10,000 × 3,600 × 12,000,000 ÷ 8 = 54 TB. CDN egress at 10x monthly catalogue plays is 540 TB.

Switching to AV1 at the same quality saves roughly 30 percent over HEVC on this kind of content: new bitrate around 8.4 Mbps, catalogue around 38 TB, egress around 380 TB. At a typical 2026 CDN price of about 0.01 dollars per GB on a tier-1 provider, that is (540 - 380) × 1,000 × 0.01 = 1,600 dollars per month, or 19,200 dollars per year saved on a relatively small library.

Switching from AV1 to AV2 at its roughly 30 percent gain takes the average to about 5.9 Mbps and monthly egress to about 265 TB: (380 - 265) × 1,000 × 0.01 ≈ 1,150 dollars more per month, but only for viewers whose devices can decode AV2, which in 2026 means software decoding. A post-VVC standard from the 2026 JVET call, if it lands near ECM's 25 percent over VVC, is a 2029–2030 decision. A neural codec trained on perceptual losses may save more, but until consumer silicon ships a neural decoder the savings are realised only in cloud transcoding, not in delivery.

The arithmetic is the punchline. Each further codec generation, on a real CDN budget, is worth roughly a thousand dollars per month per 10,000 hours of catalogue at this price, and the decoder base decides when you can collect it. That is enough to justify a codec swap every five or six years, but it is not the kind of number that funds a multi-year R&D program by itself. Neural codecs reset the math because they unlock a different axis (perception) and a different kind of saving, but they also reset the compute cost.

A pitfall to avoid: optimising the wrong objective

Engineers who already work in codecs sometimes spend optimisation budget driving down PSNR on operating points where PSNR no longer correlates with perceived quality. The classic symptom is a per-title encoding ladder that scores well on objective metrics but has waxy faces, smeared textures, and unnaturally smooth gradients on subjective review. The fix is to pick the metric (or the combination of metrics, including a perceptual one like VMAF and a learned-perceptual one like LPIPS) that matches your audience's actual judgement, and optimise for that. Otherwise the codec keeps moving down the R(D) curve on the wrong axis, and the picture looks worse at lower bitrate.

A two-by-two matrix titled Where each codec sits in 2026, placing the compute budget at the decoder against the quality objective. The columns are low compute (mobile, hardware decode) and high compute (GPU or cloud); the rows are perceptually natural and pixel accurate. Top left, Future hybrid codecs: perceptual targets inside conventional pipelines; when to ship, not yet, watch AV2 extensions and neural hybrids from 2027. Top right, Neural codecs: DCVC-FM, DCVC-RT and MS-ILLM match or beat VVC on VMAF; when to ship, cloud pre-encoding for VOD, not yet a delivery codec. Bottom left, Mainstream delivery: HEVC and AV1, broad hardware decode, low watts; when to ship, the default for OTT and live in 2026. Bottom right, Frontier conventional: VVC, AV2 and JVET ECM, best on PSNR and slow to encode; when to ship, premium tiers and broadcast. A closing line says codec choice in 2026 is a two-axis decision: the compute your decoder has, and what your audience actually judges quality by Figure 3. The same decision as a map: what your decoder can afford, and which objective your audience actually reacts to.

Frequently asked questions

What is the Shannon limit?

The Shannon limit is the theoretical maximum rate at which information can be sent over a noisy channel with arbitrarily low error, C = B · log2(1 + S/N), proved by Claude Shannon in 1948. In compression, the same name is used for the entropy floor of lossless coding and the rate-distortion floor R(D) of lossy coding.

What is the Shannon limit for compression?

For lossless compression it is the source entropy H: no code can average fewer bits per symbol without losing information. For lossy compression such as video it is the rate-distortion function R(D), the minimum bitrate for a target distortion, which depends on the content and the quality metric.

What is the Shannon curve?

In communications, the Shannon curve plots channel capacity against signal-to-noise ratio; real modulation and coding schemes sit below it. In compression, the equivalent curve is R(D), and every codec's rate-distortion curve sits above it.

How close are video codecs to the Shannon limit in 2026?

No one can say exactly, because R(D) for real video cannot be computed. The best demonstrated codec, JVET's ECM, still needs about 25–28 percent fewer bits than VVC at the same PSNR, so deployed codecs (HEVC, AV1, VVC, AV2) are not yet at the floor.

Can a codec beat the Shannon limit?

No. A codec can only beat a bound computed from a simplified model of the source. When a claim seems to break the limit, the codec is exploiting structure the model ignored, or it is being measured on a different quality metric.

Is the Shannon limit the same as entropy?

Only for lossless compression. Entropy is the floor when every bit must be reconstructed exactly; lossy video is bounded by R(D), which falls below entropy as more distortion is allowed.

What is the Shannon–Hartley theorem?

It is the channel-capacity formula C = B · log2(1 + S/N): capacity in bits per second grows linearly with bandwidth B and logarithmically with the signal-to-noise ratio. A 20 MHz channel at 30 dB SNR (S/N = 1,000) tops out near 199.3 Mbps.

Do neural codecs get closer to the Shannon limit than AV1 or VVC?

Neural codecs such as DCVC-FM and DCVC-RT beat the VVC reference encoder by about 21–22 percent in BD-rate, but independent 2024–2026 tests still place them behind ECM on PSNR. Their larger advantage is on perceptual metrics, at the cost of GPU-class decoding.

Where Fora Soft fits in

Our video and audio streaming software development team builds encoding ladders, packaging and delivery for live and VOD products, where codec-efficiency decisions show up in the line items every quarter. For catalogues, our OTT and IPTV platform development work covers the per-title encoding, multi-codec ladders (H.264, HEVC, AV1) and device matrix that decide when a new codec actually pays; for course libraries, e-learning platform development faces the same bitrate maths at lecture-video scale. The framing we apply to product teams is the same one this article ends on: separate the question "how few bits do we need for our content at our quality bar" from the question "how few bits is possible in principle". The first answer is your engineering target; the second tells you when to stop expecting more from each generation. In OTT and e-learning deployments we routinely measure first-party R(D) curves on customer content rather than rely on standardised test sets, because the wrong baseline assumption can mis-size a CDN contract by 20 percent in either direction.

Video Encoding: A Complete Guide to Video Compression - book cover

The book

Video Encoding: A Complete Guide to Video Compression

This article draws the ceiling; the book measures the distance to it. Rate-distortion theory is worked through end to end — what R(D) is for a real source, how the curve moves when you change the distortion measure, and how a BD-rate number is computed, so you can check a vendor's claim against the test conditions it came from. The codec chapters then walk the same ladder this article plots: what H.264, HEVC, AV1, VVC and AV2 each bought, where the tools ran out, and what learned compression optimises instead. Written by Nikolay Sapunov, CEO at Fora Soft.

Read it on Amazon →Kindle & paperback · Fora Soft Video Engineering Handbook

How much compression headroom is actually left on your catalogue?

Send us your current ladder and a handful of representative titles, and we will measure the rate-distortion curve on your own content: what a codec swap is worth at your quality bar, which metric your audience reacts to, and what the change does to your CDN bill and your encoder farm. You get numbers from your library rather than from a JVET test set. 250+ video products shipped since 2005.

Book a 30-min call →WhatsApp →Email us →

See our case studies


  1. Shannon, C. E. (1948). "A Mathematical Theory of Communication." Bell System Technical Journal. The founding paper of information theory; defines entropy and proves the source coding theorem. https://people.math.harvard.edu/~ctm/home/text/others/shannon/entropy/entropy.pdf. Accessed 2026-05-16. ↩

  2. Shannon's source coding theorem — Wikipedia overview of the lossless compression bound at the entropy rate. https://en.wikipedia.org/wiki/Shannon%27s_source_coding_theorem. Accessed 2026-05-16. ↩

  3. Rate-distortion theory — Wikipedia. Definition of R(D), the smallest rate at which a source can be encoded subject to a distortion constraint. https://en.wikipedia.org/wiki/Rate%E2%80%93distortion_theory. Accessed 2026-05-16. ↩

  4. Bjøntegaard, G. (2001). "Calculation of average PSNR differences between RD-curves." VCEG-M33. The Bjøntegaard delta methodology that codec teams use to compare two codecs at the same quality. https://www.itu.int/wftp3/av-arch/video-site/0104_Aus/VCEG-M33.doc. Accessed 2026-05-16. ↩

  5. Sullivan, G. J. et al. (2012). "Overview of the High Efficiency Video Coding (HEVC) Standard." IEEE Transactions on Circuits and Systems for Video Technology. Reports ~50% BD-rate savings of HEVC over H.264 on JVET test sets. https://ieeexplore.ieee.org/document/6316136. Accessed 2026-05-16. ↩

  6. Chen, Y. et al. (2020). "An Overview of Coding Tools in AV1: the First Video Codec from the Alliance for Open Media." APSIPA Transactions on Signal and Information Processing. AV1 averages ~30% BD-rate gain over HEVC on random-access configurations. https://www.cambridge.org/core/journals/apsipa-transactions-on-signal-and-information-processing/article/overview-of-coding-tools-in-av1-the-first-video-codec-from-the-alliance-for-open-media/. Accessed 2026-05-16. ↩

  7. Bross, B. et al. (2021). "Overview of the Versatile Video Coding (VVC) Standard and its Applications." IEEE Transactions on Circuits and Systems for Video Technology. VVC reports ~40-50% BD-rate gain over HEVC on UHD content. https://ieeexplore.ieee.org/document/9503377. Accessed 2026-05-16. ↩

  8. Uhrina, M., Sevcik, L., Bienik, J., Smatanova, L. (2024). "Performance Comparison of VVC, AV1, HEVC, and AVC for High Resolutions." Electronics 13(5), 953 (University of Žilina). Reports average BD-rate savings at 8K: VVC about 78%, AV1 about 63%, HEVC about 53% vs H.264. https://www.mdpi.com/2079-9292/13/5/953. Accessed 2026-10-29. ↩

  9. Alliance for Open Media (2026-06). "Alliance for Open Media Releases AV2 Codec." Press release on the AV2 v1.0 specification, libavm reference software and dav2d decoder. https://aomedia.org/press%20releases/Alliance-for-Open-Media-Releases-AV2-Codec/. Gain figures (YUV-PSNR 28.6%, VMAF 32.6% over AV1 under AOM CTC) from CNX Software, 2025-11-21: https://www.cnx-software.com/2025/11/21/aomedia-av2-open-video-codec-release-nears-delivers-around-40-bandwidth-reduction/. Accessed 2026-10-29. ↩

  10. Blau, Y., Michaeli, T. (2018). "The Perception-Distortion Tradeoff." CVPR 2018. The original paper proving that distortion and perceptual quality are mathematically at odds. https://arxiv.org/abs/1711.06077. Accessed 2026-05-16. ↩

  11. Blau, Y., Michaeli, T. (2019). "Rethinking Lossy Compression: The Rate-Distortion-Perception Tradeoff." ICML 2019. Introduces the RDP function as a generalisation of Shannon's R(D). https://arxiv.org/abs/1901.07821. Accessed 2026-05-16. ↩

  12. Li, J., Li, B., Lu, Y. (2024). "Neural Video Compression with Feature Modulation." CVPR 2024. https://arxiv.org/abs/2402.17414. The ~22% BD-rate figure vs VTM (YUV420, intra-period -1) is from Table 2 of the DCVC-RT paper (note 13). Accessed 2026-10-29. ↩

  13. Jia, Z. et al. (2025). "Towards Practical Real-Time Neural Video Compression." CVPR 2025. DCVC-RT: -21.0% BD-rate vs VTM; 1080p at 125.2 fps encode / 112.8 fps decode on A100, 118.8 / 105.3 fps on RTX 4090. https://arxiv.org/abs/2502.20762. Accessed 2026-10-29. ↩

  14. Teng, S., Jiang, Y., Gao, G., Zhang, F., Davis, T., Liu, Z., Bull, D. (2024). "Benchmarking Conventional and Learned Video Codecs with a Low-Delay Configuration." IEEE VCIP 2024. ECM saves 16.1% BD-rate over AVM and 11.0% over DCVC-FM (PSNR). https://arxiv.org/abs/2408.05042. Accessed 2026-10-29. ↩

  15. Gao, G., Feng, C., Jiang, Y. et al., Bull, D. (2026-04). "Advances in Neural Video Compression: A Review and Benchmarking." Preprints.org (not peer-reviewed). Neural codecs are competitive with VTM and AVM but trail ECM under Low Delay and Random Access. https://www.preprints.org/manuscript/202604.0035. Accessed 2026-10-29. ↩

  16. JVET Common Test Conditions and AOMedia Common Test Conditions — the standardised test sets and configurations used to make BD-rate comparisons meaningful. https://www.itu.int/en/ITU-T/studygroups/2017-2020/16/Pages/video/jvet.aspx. Accessed 2026-05-16. ↩

  17. ITU-T SG21 / JVET page (2026): joint call for proposals on video compression with capability beyond VVC issued July 2026. https://www.itu.int/en/ITU-T/studygroups/2025-2028/21/video/Pages/jvet.aspx. ECM version 18 at ~28% over VVC and the January 2027 / end-2029 timeline: Pennington, A., "AI Helps InterDigital Reach Beyond VVC in Race to Develop Next Gen Codec," Streaming Media, 2025-10-17, https://www.streamingmedia.com/Articles/News/Online-Video-News/AI-Helps-InterDigital-Reach-Beyond-VVC-in-Race-to-Develop-Next-Gen-Codec-171971.aspx. Accessed 2026-10-29. ↩