Royalty-free AV2 video codec with 30% bitrate savings
We return to the AV2 news, the next open video codec from the Alliance for Open Media, which is nearing completion after five years of development. The specification is expected by the end of 2025, with all core tools finalized and high-level syntax work in progress. In current tests, AV2 shows around 30% lower bitrate than AV1 at the same quality.
Andrey Norkin, who works at Netflix as research scientist and co-chairs at AOM, has confirmed that recent results show a 28.63% bitrate reduction compared to AV1 in PSRN-YUV (14:1:1) results and 32.59% in VMAF, the latter stands for Video Multi-Method Assessment Fusion, which is a video quality metric developed by Netflix.
If you expect some groundbreaking AI algorithm that does the miracle job here, it’s not that kind of codec. The AOM group does mention possible AI extensions being used, but generally this is simply math, algorithms, and just pure creative thinking of people working on such projects.
Here are the most important slides:
Source: AOM/Andrey Norkin
From what I was able to hear, the AV2 codec keeps the hybrid block-based structure familiar from AV1 but introduces larger 256×256 superblocks, fully recursive partitioning, and smarter separation of luma and chroma splits. Prediction has been strengthened with new data-driven intra modes, improved chroma-from-luma modeling, and a ranked reference system that selects from up to seven past frames. Inter prediction adds temporal interpolation (TIP), improving motion handling in high-resolution or fast-moving scenes.
AV2 introduces a unified exponential quantizer with wider range and more precision for 8-, 10-, and 12-bit video. Trellis-coded quantization (an improved method of quantizing) and user-defined matrices improve control at lower bitrates. The transform system now uses learned and cross-component transforms to retain texture with fewer artifacts, while coefficient coding was refined for screen and mixed-content material.
AV1 deblocking vs AV2 deblocking, Source: AOM/Andrey Norkin
Filtering and post-processing were also upgraded. Now, a single generalized deblocker preserves detail better (as shown above) and new filters such as the Guided Detail Filter and Cross-Component Sample Offset further clean up compression noise. Film grain synthesis is more flexible, and the codec supports extended multi-layer and stereo video designs. All AV2 tools have been validated for hardware efficiency, and the group’s next focus is on encoder optimization and possible extensions for higher bit depths and AI-assisted profiles.
I extracted the entire presentation for you to check, but keep in mind some slides are in the wrong positions.
Source:
About the Author
