Core Technologies from VP9, Daala, and Thor in AV1

The AOMedia Video 1 (AV1) codec is an open-source, royalty-free video standard developed by the Alliance for Open Media (AOMedia) to deliver superior compression efficiency for web and streaming media. Designed to succeed VP9, AV1 is a collaborative synthesis built upon the foundational architectures and research of three predecessor projects: Google’s VP9, Xiph.Org/Mozilla’s Daala, and Cisco’s Thor. This article outlines the specific core technologies that AV1 adopted from each of these three projects and how they shaped the final standard.

Core Technologies from VP9

Google's VP9 served as the baseline pipeline and bitstream framework for AV1. Rather than designing a codec from scratch, AOMedia used VP9's codebase as the starting point, expanding its capabilities across several domains:

• Baseline Bitstream Architecture: AV1 preserves VP9’s high-level frame structure, tile division, and reference frame buffer management, providing a familiar framework for hardware implementation.
• Partitioning Framework: VP9 introduced the concept of recursive four-way block splitting starting from 64x64 superblocks. AV1 expanded this structure to support 128x128 superblocks and added rectangular, non-square splits (such as 1:2, 2:1, 1:4, and 4:1) along with T-shaped partitions.
• Deblocking Filter: The foundational in-loop deblocking filter in AV1 directly descends from VP9's edge-filtering system, smoothing artifact boundaries across block edges prior to newer filtering stages.
• Inter and Intra Prediction Base: The core motion vector modeling and standard directional intra-prediction angles originated in VP9, which AV1 extended with finer angle gradients and multi-parameter affine transforms.

Core Technologies from Daala

The Daala project, led by Xiph.Org and Mozilla, focused on experimental, non-traditional compression methods. While AV1 did not adopt Daala’s lapped transforms entirely, several of Daala's key innovations were integrated:

• Multi-Symbol Entropy Coding: VP9 used a binary arithmetic coder that processed one bit per cycle. Daala contributed an advanced multi-symbol arithmetic coder capable of processing up to 8-bit symbols directly, significantly boosting throughput and facilitating parallel hardware decoding.
• Chroma-from-Luma (CfL) Prediction: Daala pioneered the cross-component prediction technique known as Chroma-from-Luma. CfL models chroma pixels as a linear function of reconstructed luma pixels, exploiting the strong correlation between color and luminance channels to drastically reduce chroma residual data.
• Directional Deringing: Daala’s directional deringing filter analyzes edge directions within a block and applies smoothing along—rather than across—those edges. This technique was combined with Thor’s filtering research to create AV1's hybrid post-processing stages.

Core Technologies from Thor

Developed by Cisco to provide a royalty-free video standard optimized for real-time communications, Thor contributed critical motion prediction tools and filtering enhancements:

• Constrained Low-Pass Filter (CLPF): Thor utilized a selective smoothing filter designed to reduce ringing and mosquito noise while preserving sharp edge details.
• Constrained Directional Enhancement Filter (CDEF): AV1's signature in-loop filter, CDEF, is a direct merger of Thor’s CLPF and Daala’s directional deringing filter. CDEF first identifies the dominant edge direction of an 8x8 block and applies a non-linear low-pass filter aligned with that direction, suppressing artifacts without blurring fine lines.
• Advanced Motion Vector Referencing: Thor contributed spatial and temporal motion vector prediction techniques, allowing AV1 to build sophisticated candidate lists from neighboring blocks and past reference frames to improve inter-frame motion accuracy.

Conclusion

By using VP9 as a stable structural foundation, integrating Daala's advances in entropy coding and cross-channel prediction, and incorporating Thor’s motion modeling and filtering expertise, AV1 achieved a compression performance exceeding that of its predecessors while maintaining a royalty-free deployment path.