The heart of HEVC’s complexity lies in how it decides to chop each picture into blocks. The standard divides every frame into large coding tree units, which are then recursively split into smaller coding units in a quadtree structure. For each coding unit, the encoder must evaluate a set of prediction unit, or PU, partition modes, including symmetric square partitions of 2N by 2N and N by N, asymmetric partitions, and special modes such as skip and merge. For every candidate, the encoder runs motion estimation, motion compensation, and a rate-distortion optimization loop that weighs the number of bits spent against the distortion introduced. Because the quadtree can nest several levels deep and every node spawns multiple mode candidates, the number of decisions an encoder must make per frame explodes combinatorially. This recursive block partitioning and the associated PU mode decision are, as the authors note, the main factors behind HEVC’s heavy computational load.

The insight behind the new work is that many of these exhaustive evaluations are wasted effort. In real video content, a large fraction of coding units end up choosing the simplest option anyway: the 2N square PU partition, which treats the whole coding unit as a single prediction block. This is especially true for the skip mode, in which no residual data is coded at all, and for merge mode, in which motion information is simply copied from a neighboring block. When the encoder nevertheless tests all the non-square and asymmetric alternatives, it burns cycles on candidates that almost never win. If a smart rule could predict, early in the process, that the square mode will turn out to be optimal, the encoder could skip the expensive evaluations entirely and move on to the next coding unit.

Building such a rule is harder than it sounds, because the cost of a given mode depends on the content of the scene. A static talking head favors big square blocks and skip decisions, while a fast pan across a crowded stadium favors fine partitions with asymmetric shapes. Fixed thresholds tuned on one type of content fail badly on another. The researchers’ answer is an online threshold model, one that adapts to the sequence as it is being encoded rather than relying on constants learned offline. The model is built from statistical analysis of the rate-distortion costs of the best PU partition modes of previously encoded coding units, so it continuously tracks the cost distribution that the current sequence is actually producing.

The mechanism works as follows. As the encoder processes coding units, it records the rate-distortion cost of the winning partition mode for each one. From this running history, the scheme establishes a threshold that characterizes when a 2N square PU partition is likely to be optimal for the unit currently under evaluation. When the cost of the square mode for the current coding unit falls below that threshold, the encoder concludes that the square partition is the right answer and skips the evaluation of all non-square PU partition modes entirely. This covers the full family of square decisions, including both the merging mode and the advanced motion vector prediction, or AMVP, mode, as well as the skip mode. If the cost does not fall below the threshold, the encoder falls back on the standard exhaustive search, so the scheme never gambles more than the evidence supports.

The word online in the model’s name is doing important work. Because the threshold is derived from the costs of previously encoded units within the same encoding session, it automatically adjusts as the character of the video changes. A scene cut from a still interview to a high-motion action sequence shifts the cost statistics, and the threshold follows. This adaptivity is precisely what separates the approach from earlier fast algorithms that relied on precomputed parameters, spatial features, or offline-trained classifiers. Those methods can perform well on the content they were tuned for but degrade when the encoder meets something new. An online model, by contrast, learns from the very stream it is compressing, which makes it robust across sequences with very different motion characteristics.

The new scheme sits within a rich research tradition. Since HEVC was finalized, researchers have attacked its complexity from many angles: fast coding unit size decision algorithms using texture and spatial features, data mining approaches that train decision trees on encoding statistics, Bayesian adaptive methods that model the probability of partition outcomes, and, more recently, machine learning techniques including convolutional and LSTM neural networks that predict partition structures directly. The reference list of the paper spans more than a decade of this effort, from early intra-coding speedups through recent work on fast partitioning for HEVC-to-VVC transcoding and deep learning frameworks for coding tree unit partitioning. What distinguishes the present contribution is its focus on the inter-prediction PU mode decision stage and its deliberately lightweight, self-updating statistical formulation, which avoids the training pipelines and memory footprints that neural approaches require.

The experimental results reported by the authors show that the proposed method outperforms previous fast schemes, with negligible degradation in coding efficiency and video quality compared with the HEVC test model, known as HM, the reference software that defines the standard’s performance baseline. In the video compression community, negligible loss against HM is the gold standard for a fast algorithm: it means the encoder saves enormous computation while producing bitstreams whose size and decoded quality are essentially unchanged. Notably, the advantage held up across test sequences with different motion characteristics, which is exactly the scenario where fixed-threshold and offline-trained methods tend to falter. The authors attribute this robustness to the online nature of the threshold, which keeps the skip decisions aligned with the actual cost statistics of the content at hand.

The practical implications reach well beyond academic benchmarks. Every hour of video encoded for a streaming service consumes real energy and real money in data center CPU time, and HEVC encoding is among the most expensive operations those facilities perform. A scheme that eliminates a large share of the mode decision workload translates directly into higher throughput per server, lower latency for live encoding pipelines, and reduced power consumption, which matters both for cost and for the growing scrutiny of the streaming industry’s carbon footprint. On the other side of the pipe, the same principles ease the burden on devices that transcode or re-encode video locally, from broadcast head-ends to mobile handsets capturing and uploading footage. Because the method preserves rate-distortion performance, these gains come without the visible quality penalties that aggressive shortcut encoders often introduce.

The work also offers a lesson about where compression research is heading. As the successor standard H.266/VVC pushes complexity even higher with its larger coding tree units and richer partitioning toolbox, the demand for intelligent early-decision mechanisms only intensifies, and several of the works cited in the paper already apply similar ideas to VVC. The Pyongyang team’s online threshold model demonstrates that careful statistical modeling of the encoder’s own behavior, updated in real time, can rival far heavier machinery. It is a reminder that in video coding, as in many computational fields, the fastest path forward is often not to compute harder but to decide sooner which computations are not worth doing at all. For the billions of hours of video flowing across networks every year, that difference between computing harder and deciding sooner adds up to a very large number indeed.

Subject of Research: Fast prediction unit mode decision in HEVC video encoding using an adaptive online threshold model

Article Title: A research of online-threshold model for 2 N square PU- early detection in HEVC

Article References: Mun, K. S., Huang, J. G., & Jong, H. M. (2026). A research of online-threshold model for 2 N square PU- early detection in HEVC. Multimedia Tools and Applications, 85(9), Article 734. https://doi.org/10.1007/s11042-026-21906-2

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21906-2

Keywords: HEVC, H.265, video compression, prediction unit, rate-distortion optimization, online threshold model, encoding complexity, skip mode, merge mode, AMVP, fast algorithms, Multimedia Tools and Applications

The conceptual foundation of the new framework departs sharply from the dominant view of video compression that has prevailed for decades. Classical communication theory, rooted in the Shannon-Weaver model articulated by Claude Shannon in 1948, distinguishes three levels of communication. Level A concerns the technical problem of transmitting data accurately, Level B addresses whether transmitted symbols convey the intended meaning, and Level C concerns the effectiveness problem: whether the received information produces the desired outcome. Video standards such as HEVC have concentrated almost exclusively on Level A, maximizing signal fidelity under constrained bandwidth by minimizing distortion between the original and the reconstructed signal. GVC instead places Level C at the heart of its design, asking not whether every pixel survives the journey but whether the reconstructed video meets perceptual expectations or supports the task at hand.

The driving principle behind GVC is elegantly simple to state: trade computation for compression rate. Instead of transmitting detailed visual data, the framework encodes video into extremely compact representations and delegates content reconstruction to the receiver, where powerful generative priors synthesize high-quality video from minimal transmitted information. The researchers offer a vivid metaphor to explain the shift. Traditional compression is like photographing a painting and sending the photograph; GVC is like describing the painting’s composition and style, then relying on an AI painter at the far end to recreate it. Modern generative video models are so expressive that they can synthesize convincing footage from sparse latent representations, or in the limit even from pure noise guided by learned priors. That capability transforms the encoder’s job from preserving every pixel to selecting and transmitting only the most task-relevant information.

The system is built from two primary components working in tandem. On the sending side, a neural encoder, a pre-trained neural network, ingests an input video sequence, which might be surveillance footage, a video call stream, or a live broadcast, and compresses it into a set of compact representations called compressed tokens. These tokens blend discrete and continuous elements: compressed keyframes, high-level descriptors of video segments, and low-level continuous features that together capture the essential semantics and motion dynamics of the scene while drastically reducing dimensionality. The tokens are further encoded into a bitstream using techniques such as residual coding to squeeze out remaining redundancy. On the receiving side, a pre-trained diffusion-based generative video model performs what is essentially a conditional video generation task. Some tokens serve as direct inputs to the denoising process while others act as conditioning signals, and the model synthesizes frames that are visually faithful to the original input.

What gets transmitted depends, critically, on the purpose of the reconstruction. If the goal is human perception, the encoder sends features that help the generative decoder produce perceptually similar content. If the goal is machine understanding, for instance segmentation or recognition by a downstream algorithm, the encoder focuses on semantically meaningful representations instead. This task-oriented orientation is where GVC aligns itself with the AI Flow framework, proposed by TeleAI at the end of 2024, which envisions communication networks distributing intelligence for ubiquitous AI-powered services. The theoretical underpinning draws on the concept of Information Capacity, a measure of how efficiently generative models compress data, as well as earlier work on task-oriented feature compression for multimodal understanding via device-edge co-inference. The GVC concept itself was first introduced publicly by TeleAI at the World Artificial Intelligence Conference in mid-2025, where a prototype for maritime communications demonstrated ultra-low bitrate video transmission over bandwidth-limited satellite links.

Extreme compression, however, introduces a new bottleneck: the computational cost of high-quality generative reconstruction. Diffusion-based decoders are computationally intensive, and hardware, power, and latency constraints impose an upper bound on how much computation can realistically be traded for compression, especially in real-time applications like video conferencing or edge-device streaming. The team’s answer is a second, complementary principle: trading compression rate for practicality. By sacrificing a small fraction of the compression ratio, the system can send richer latent representations that reduce reliance on massive generative models, unlocking the use of smaller and faster decoders. The researchers further apply model compression techniques to shrink key components such as 3D variational autoencoders, and they employ distillation and sampling acceleration methods for the diffusion-based decoder to lower inference time. The result is a flexible balance across the compression-computation-quality triangle that adapts to whatever resources the deployment environment offers.

The empirical results are striking. Benchmarked on the standard MCL-JCV dataset using a 14-billion-parameter video generative model, GVC maintained competitively high perceptual quality, measured with the Learned Perceptual Image Patch Similarity metric, at an average bitrate of just 0.008 bits per pixel, equivalent to roughly a 0.02% compression rate. Conventional video coding schemes exhibit a substantial performance gap at this bitrate; on certain challenging sequences, traditional methods require approximately six times more bandwidth to match the perceptual quality achieved by GVC. In one benchmark example, the framework achieved visually compelling reconstruction at 0.005 bits per pixel. In real-world surveillance scenarios, the compression can go further still: the team reports bitrates below 0.002 bits per pixel, crossing the 0.01% compression threshold, while retaining sufficient visual quality for the task.

Importantly, the extreme compression does not come at the expense of semantic integrity. To test downstream utility, the researchers applied the reconstructed videos to video object segmentation on the DAVIS2017 benchmark, evaluating performance with the Jaccard index, contour accuracy, their average, and contour recall. The compressed-and-regenerated videos achieved highly competitive segmentation results, indicating that even at astonishingly low bitrates the framework preserves the semantic information machines need to understand a scene. This validates the core promise of effectiveness-level communication: the reconstructed video is not merely visually plausible but genuinely useful for the tasks the transmission was intended to serve.

Deployment readiness was demonstrated on real hardware. After miniaturization, distillation, and quantization of the generative decoder, the system can reconstruct a group of 29 frames in a single pass with inference latency of around two seconds on consumer-grade GPUs, a response time comparable to what users routinely experience with large language models. Although the miniaturized model incurs some loss in visual quality and bandwidth efficiency relative to its full-scale counterpart, it still maintains competitively high perceptual quality, with a demonstrated LPIPS score of 0.273 on a sample sequence. That combination of speed, quality, and modest hardware requirements makes GVC a plausible candidate for the environments that need it most: emergency rescue operations, remote surveillance, narrowband mobile networks, in-vehicle and wearable devices, and maritime satellite links where bandwidth is scarce and expensive.

The authors frame GVC not merely as another codec but as a task-oriented communication paradigm tailored for the era of generative intelligence. By transmitting only what is necessary for perception and decision making, and letting generative priors at the receiver do the heavy lifting of reconstruction, the framework opens the door to communication systems that are more efficient, adaptive, and intelligent than the fidelity-obsessed pipelines of the past. Whether video transmission at one hundredth of one percent of its original size becomes a routine capability will depend on further advances in generative model efficiency and edge computing, but this work offers a credible, empirically validated path toward that frontier, and a glimpse of a future in which the networks we build carry descriptions rather than copies, and understanding rather than pixels.

The work appears as a brief communication in Vicinagearth, an open-access journal, published on 23 March 2026 as volume 3, article number 7, with a correction issued on 3 June 2026. Its placement at the intersection of coding and information theory, computer vision, and multimedia systems reflects the increasingly hybrid nature of compression research, where ideas from generative modeling are being grafted onto classical transmission problems.

Historically, the Shannon-Weaver model dates to the 1940s, and the authors note that video communication technology has spent decades optimizing its Level A, the technical problem of accurate signal delivery. The Information Capacity metric, proposed as a way to evaluate how effectively generative models compress data, laid methodological groundwork for this line of research, and in early 2025 the same group extended the approach to task-oriented communications for multimodal understanding via device-edge co-inference. The maritime prototype unveiled at the World Artificial Intelligence Conference demonstrated ultra-low bitrate transmission over bandwidth-limited satellite connections, a setting where every saved bit carries direct operational value.

Within the AI Flow framework, the researchers position GVC as opening new possibilities for video communication in bandwidth- and resource-constrained environments such as emergency rescue, remote surveillance, and mobile edge computing, describing it as a viable path toward an effective, efficient, scalable, and practical video communication paradigm.

Subject of Research: Extreme low-bitrate video compression using generative AI models to reconstruct video from minimal transmitted information

Article Title: Generative video compression: towards 0.01% compression rate for video transmission

Article References: Generative video compression: towards 0.01% compression rate for video transmission. (n.d.). https://doi.org/10.1007/s44336-026-00035-2

Image Credits: AI Generated

DOI: 10.1007/s44336-026-00035-2

Keywords: generative video compression, video compression, task-oriented communication, AI Flow, diffusion models, bitrate, Shannon-Weaver model, edge computing, video transmission, LPIPS, surveillance, neural encoder