Streaming Learning Center Where Streaming Professionals Learn to Excel
Related Articles
AI video compression standards have moved from the research lab to the standards committee. This post summarizes the current status of formal standardization efforts as of July 2025, the groups driving them, and the anticipated timeline for real-world deployment. If you’re wondering whether the next generation of video codecs will be AI-native or just AI-enhanced, this should give you a grounded answer.
Notably excluded are works outside of standards bodies, such as those performed by Deep Render, which you can read about here and here. Working outside the structure and collective nature of standards-setting bodies, Deep Render appears to be the closest to an actual working AI-based codec.
Contents
MPEG, ITU, AOM, and MPAI: Who’s Doing What
MPEG continues to push forward on several fronts. Its MPEG‑AI standard (ISO/IEC 23888) includes parts covering video coding for machines, feature coding for machines, and AI‑based point cloud coding. The Video Coding for Machines (VCM) initiative targets scenarios like surveillance, autonomous vehicles, and industrial use cases—where machine analysis, not visual quality, is the priority. More on this below.
Its Enhanced Compression Model (ECM) project has reached version 15, demonstrating roughly 25% bitrate savings over VVC in random-access configurations and up to 40% for screen-content sequences. ECM is a likely foundation for the future H.267 codec, which aims for a ~40% bitrate reduction relative to VVC. At the same time, MPEG is continuing its exploratory work on Neural Network‑Based Video Compression (NNVC), now up to version 7 of its algorithm and software specification.
JVET (a collaboration between MPEG and ITU) is also laying the groundwork for H.267. In Geneva on 17 January 2025, ITU-T and ISO/IEC held a joint workshop on “Future video coding – advanced signal processing, AI and standards,” making clear that next-gen codecs will be evaluated not just on compression efficiency but also latency, deployability, and use‑case relevance. The H.267 standard is currently projected to be finalized between July and October 2028, with meaningful deployment likely not occurring until around 2034–2036.
AOMedia is sticking with AV1, while layering in AI-based enhancements like encoder tuning and scene analysis. No AI-native codec efforts here yet. The group is also working on IAMF (Immersive Audio Model and Formats), which could benefit from AI but isn’t video-focused.
MPAI is the most aggressive on the AI-native front. Its MPAI‑EVC project focuses on AI‑enhanced modules within traditional pipelines, while its AI-based End-to-End Video Coding (MPAI‑EEV) is a clean‑slate, neural‑only effort now on version 4 of its reference model. Both are available as open source.
What’s Working, What’s Holding Things Back
Recent research claims strong results: up 22.7% better compression than H.266 using advanced neural codecs. There’s also progress in context modeling and post-processing filters, all of which could end up in future standards—if they prove practical.
But practicality is the bottleneck. Samsung says network quality has made bitrate less of a concern; what matters now is low decoder complexity, especially on mobile. Amazon wants features like film grain synthesis and post-standardization update paths. MainConcept reminded everyone that standardization doesn’t equal adoption. H.267 may hit the spec sheet in 2028 but some don’t expect real deployment closer to 2035–2040.
The table below summarizes the current standardization efforts as of July 202. None of the AI-native codecs appears near readiness for commercial use. Even the enhanced pipelines are still being prototyped.
| Organization | Standard/Initiative | Focus/Objective | Latest Stage (July 2025) | Leading Group(s) |
|---|---|---|---|---|
| MPEG | MPEG-AI: Video Coding for Machines | Machine-optimized video compression | Committee Draft (CD) | MPEG WG 4 |
| MPEG | MPEG-AI: Feature Coding for Machines | Feature-based compression for machine analysis | Ongoing development | MPEG WG 4 |
| MPEG/ITU | HEVC/AVC AI-Enhanced SEI | Neural network post-filtering, AI metadata | CDAM (HEVC: Oct 2025, AVC: 2026) | MPEG WG 5, ITU-T SG 16 (JVET) |
| MPEG | MPEG Immersive Video (MIV) | Volumetric/immersive video coding | Final Draft International Standard | MPEG WG 4 |
| AOM | AVM (AOMedia Video Model) | Next-gen, AI-enhanced, royalty-free codec | Public testing, pre-standard | AOM Codec Working Group (CWG) |
| AOM | Volumetric Visual Media (VVM) | Immersive/volumetric video coding | Working group activity | AOM VVM Working Group |
| ITU | AI-Enabled Multimedia Coding | AI in video/audio coding, quality assessment | Multiple standards; ongoing | ITU-T SG 16 (VCEG), Q5/16 |
| ITU/ISO/IEC | Multimedia Authenticity (AI) | AI watermarking, deepfake detection | New collaboration, early stage | ITU, ISO, IEC |
AI Codecs: Not Just for Humans
Notably, some AI-based codecs aren’t being built for human viewing at all. VCM (Video Coding for Machines), for example, is optimized for tasks like object detection or scene classification in surveillance, robotics, or autonomous vehicles. FCM (Feature Coding for Machines) goes a step further; instead of compressing pixels, it focuses on encoding extracted features such as bounding boxes or object vectors. These use cases demand new metrics and architectures, and while they share tools with traditional codecs, the end goals are fundamentally different.
On the hardware side, AI codecs might finally have a path to deployment thanks to growing support for NPUs in mobile devices. However, while FCM appears to emphasize NPU usage, broader industry adoption is lagging. I’ve questioned why NPUs aren’t being mandated for H.267, noting that NPU capability hasn’t yet been prioritized by industry stakeholders. Whether the industry will adopt NPUs at scale and whether that’s enough to make AI-native codecs viable on phones or TVs remains to be seen.
Conclusion
The key takeaway is that standardization is happening, but full AI-native codecs aren’t ready and likely won’t be for another decade. Meanwhile, AI-enhanced tools are finding their way into traditional encoders and codec pipelines. If you’re evaluating next-gen codecs, don’t expect AI to save bandwidth overnight. But it is reshaping the design criteria—and deployment decisions—for what comes after H.266.
