Neural Data Compression: Can AI Replace Traditional Codecs?
Data compression is one of the technologies that makes modern computing practical. Photos, streamed films, video calls, music libraries, cloud backups and software downloads would require far more storage and bandwidth without codecs that reduce the amount of data involved. For decades, these codecs have relied mainly on carefully designed mathematical rules and signal-processing techniques. Artificial intelligence is now changing that approach. Instead of engineers defining every stage of compression by hand, neural systems can learn how to represent images, sound or video efficiently from large collections of examples. By 2026, neural compression has moved well beyond laboratory demonstrations, particularly in image and audio coding. Yet replacing familiar codecs such as JPEG, AV1, AV2, VVC, AAC or Opus is a much bigger challenge than simply producing a smaller file. Compatibility, processing requirements, reliability and long-term support are equally important.
How Neural Compression Changes the Job of a Codec
A traditional codec normally follows a sequence of techniques developed specifically for a certain type of information. An image codec might separate an image into blocks, transform colour and brightness information, remove details that people are unlikely to notice and then store the remaining values efficiently. Video codecs add another major trick: instead of saving every frame independently, they predict parts of one frame from earlier or later frames. Audio codecs use similar ideas based on what the human ear is more or less likely to hear. Modern codecs have become extremely sophisticated, but much of their behaviour still comes from rules designed and refined by engineers.
Neural compression changes where those rules come from. A neural network is trained on many examples and learns a compact internal representation of the original material. For an image, that representation can retain information about shapes, textures, edges and colours while discarding less important detail. A decoder then uses the compressed representation to reconstruct an image that is intended to look close to the original. The same general idea can be applied to speech, music and video. Engineers still design the overall system and decide how it should be trained, but the model learns many of the decisions that traditional codecs handle through fixed algorithms.
This makes neural compression particularly interesting when the goal is perceptual quality rather than perfect numerical similarity. Two images can differ at the individual-pixel level while appearing almost identical to a person. A learned codec can be trained to concentrate its limited number of bits on characteristics that matter more to human vision. The same principle works with sound, where maintaining natural speech or recognisable musical detail may be more valuable than reproducing every sample exactly. This should not be confused with lossless compression, however. When a document, program, database or executable file must be recovered byte for byte, there is no acceptable room for an approximate reconstruction.
Where AI Compression Is Already Proving Useful
Still-image compression provides one of the clearest signs that neural coding is becoming a practical technology. ISO/IEC 6048-1:2025, better known as JPEG AI, was published in 2025 as the first international image coding standard built around an end-to-end learning-based approach. Additional JPEG AI specifications covering profiling and the file format followed in 2026. This matters because a successful codec requires more than a good research model. Different applications must be able to create, exchange and decode compatible files. JPEG AI provides a defined bitstream and decoding process rather than leaving each developer with an incompatible neural compression system.
Research continues to show that learned image compression can move beyond the efficiency of conventional designs. In May 2026, Apple researchers presented a practical learned image codec designed around both perceptual quality and realistic device performance. Their reported subjective tests showed substantial bitrate savings compared with several advanced alternatives, including AV1, AV2, VVC, JPEG AI and experimental compression technology. The researchers also demonstrated 12-megapixel image encoding and decoding directly on a current smartphone rather than relying only on powerful data-centre hardware. It is still research rather than evidence that every image format should immediately be replaced, but it demonstrates how quickly the processing gap between neural and conventional codecs is narrowing.
Audio is another strong candidate. Neural audio systems such as Google SoundStream and Meta EnCodec learn compact representations of speech, music and other sounds. Earlier neural speech codecs were especially attractive at extremely low bitrates, where conventional codecs often produce obvious robotic or metallic artefacts. Newer systems are designed for a wider range of audio and can preserve convincing sound using surprisingly little data. Neural audio codecs have also become important in generative AI because their compressed representations can act as audio tokens that other models process. Even so, widely supported codecs such as Opus remain difficult to displace in everyday calls, streaming and browser communication because their implementations are mature, efficient and available across an enormous range of devices.
Why Traditional Codecs Are Still Difficult to Displace
The biggest advantage held by established codecs is not always compression efficiency. It is compatibility. A successful media format may be supported by operating systems, browsers, graphics processors, televisions, cameras, editing software, game consoles and mobile devices. Dedicated hardware can decode popular video formats using far less power than a general processor performing the same work. Replacing such an ecosystem requires manufacturers, software developers and content providers to support the new codec at roughly the same time. A neural codec that produces a file 20 or 30 per cent smaller can still be impractical if most users cannot open that file without special software.
Processing cost is another important limitation. Neural compression requires mathematical operations that may run efficiently on a modern neural processing unit but much less efficiently on older hardware. A powerful model may save network bandwidth while consuming more electricity during encoding or decoding. That trade-off matters in smartphones, cameras, battery-powered devices and large streaming services handling millions of files. Model size also matters because the decoder itself may require additional memory or stored neural-network parameters. Improvements in mobile AI hardware are steadily reducing these problems, but a comparison based solely on final file size leaves out an important part of the real cost.
Traditional codecs also benefit from decades of testing. Their behaviour is understood across unusual images, damaged files, weak network connections and different hardware implementations. For archives, broadcasters and professional production systems, predictable decoding years or even decades later can be more valuable than saving a few additional megabytes today. Neural codecs therefore need clearly specified models, stable decoder behaviour and reliable version management. JPEG AI is important partly because standardisation addresses some of these concerns. Individual research codecs may achieve exceptional results, but a model stored in a research repository is not automatically suitable for long-term preservation or mass distribution.
The Difference Between Smaller Files and Better Compression
Claims about one codec being better than another require careful interpretation. The smallest file is not necessarily the best result. Compression tests normally compare bitrate with reconstructed quality, but quality itself can be measured in several ways. A mathematical metric may reward an image that is numerically close to the source even when another image looks better to human observers. Neural systems can be trained for perceptual quality and therefore perform especially well in listening or viewing tests. Conventional codecs can be optimised for similar goals too. Fair comparisons should consider the same source material, similar operating conditions, encoding time, decoding time and subjective quality rather than relying on a single headline percentage.
There is also an important boundary between preserving information and creating a convincing approximation. Some experimental generative compression methods can reconstruct plausible detail that was not transmitted explicitly. At very low bitrates, this may produce a more natural-looking face, landscape or texture than a conventional codec covered in blocks and blur. That is useful for some entertainment and communication tasks, but it can be unacceptable where every visible detail has evidential value. Security footage, scientific imagery, medical material and technical inspection data may require preservation of specific features rather than an image that merely looks realistic. A visually convincing reconstruction is not automatically a faithful record.
The case for AI replacement becomes weaker still with general-purpose lossless data. Compression tools used for source code, software packages, documents, databases and backups must reproduce exactly the same sequence of bytes after decompression. Machine learning can estimate patterns in such data and may contribute useful ideas to compression research, but established methods are fast, predictable and relatively inexpensive to run. They also work without requiring a large trained model matched to a particular type of content. For ordinary PC storage and file archiving, neural compression is therefore more likely to complement established lossless techniques than suddenly make familiar archive and compression algorithms obsolete.
What the Compression Landscape Looks Like in 2026
Neural compression is advancing at the same time as conventional codec development, so AI is competing against a moving target. The ITU approved an updated fourth edition of H.266 Versatile Video Coding in January 2026. AOMedia also completed version 1.0 of the AV2 specification in May 2026. AV2 follows AV1 and targets higher compression efficiency together with newer uses including advanced screen content and immersive media. These developments show that traditional codec engineering has not reached a dead end. Engineers continue to find better prediction methods, transforms, coding tools and implementation strategies without replacing the entire codec with a neural network.
At the same time, the organisations responsible for future video standards are clearly taking AI seriously. At its July 2026 meeting, the Joint Video Experts Team issued a call for technologies for a generation of video compression beyond VVC. The process explicitly allows both conventional signal-processing methods and neural-network or other AI-based technologies to compete. Formal evaluation of submissions is planned to begin in 2027, with a tentative target of completing a first next-generation standard around the end of 2029. As a result, there is no established all-neural successor to today’s major video codecs in 2026, but neural techniques are being evaluated as potential components of whatever comes next.
A hybrid direction may prove more realistic than a clean divide between AI codecs and traditional codecs. Neural networks can improve selected tasks while an established codec structure continues to handle the bitstream, timing, compatibility and basic decoding process. Machine learning can assist prediction, filtering, image restoration, rate control or content analysis without requiring every part of the codec to become neural. Modern processors increasingly include dedicated hardware for machine-learning workloads, making these combinations more practical. This gradual approach gives developers a way to obtain some of the efficiency or quality benefits of AI while retaining more of the interoperability and predictable behaviour associated with existing formats.
Can AI Replace Traditional Codecs?
In individual applications, the answer is already becoming yes. Learning-based image coding has reached international standardisation through JPEG AI, and neural audio codecs have demonstrated impressive results in low-bitrate speech and general audio. Specialised services that control both encoding and decoding can adopt a neural system without waiting for universal device support. The wider answer is different. Video streaming, web media, broadcasting, local storage and lossless computing all have different requirements. No single neural method currently provides a compelling replacement for every conventional codec across all of these tasks.
The more likely change is that the definition of a codec will evolve. Future codecs may contain conventional components, neural components and models trained for particular reconstruction tasks, while users simply see a supported file or media stream. In other cases, a fully learned codec may use a standardised decoder so that files remain interoperable between different applications. The boundary between compression and AI processing may become increasingly difficult to see. What matters to users is not whether a neural network sits inside the decoder, but whether the result offers acceptable quality, speed, battery use, storage efficiency and reliable playback.
By 2026, neural compression should therefore be treated as a genuine new branch of codec design rather than a fashionable replacement label. It has already produced an international image standard, practical audio systems and research results that challenge some of the best conventional methods. At the same time, AV2, VVC and established lossless compressors demonstrate why traditional techniques will remain relevant for years. Anyone choosing a codec should consider compatibility, bitrate, visual or audio quality, processing cost, latency, licensing, long-term support and the need for exact reconstruction. AI can replace traditional codecs where its advantages justify those trade-offs, but the broader future of compression is much more likely to involve coexistence and combination than a complete takeover.
