AV1 (AOMedia Video 1) is a free next-generation video compression codec developed by the Alliance for Open Media consortium. In simple terms, the technology allows transmitting images in 4K resolution and higher at half the bitrate compared to outdated standards, without losing visual quality and without requiring royalty payments.
The codec is being implemented on streaming platforms like YouTube and Netflix, where traffic savings are critically important when broadcasting millions of hours of content. Its support is implemented in Chromium-based browsers, in NVIDIA, AMD, and Intel GPUs, as well as in Android and iOS operating systems for media file playback. The technology is used in video conferencing applications (Google Meet) and in live streaming services where sub-second latency is required.
The main challenge with AV1 is the extremely high computational load during software encoding, which can be hundreds of times slower than real-time without a hardware accelerator. This creates problems for content creators with limited CPU resources. Hardware decoding on older devices is impossible, causing rapid battery drain and heating during software playback. There are also still delays in widespread support on budget TVs and set-top boxes.
How AV1 works
The operating principle of AV1 is based on a hybrid compression scheme with a radical expansion of prediction tools. Unlike the H.264 codec, where block partitioning was rigidly fixed, AV1 introduces a tree-structured superblock system up to 128×128 pixels in size, capable of dynamically splitting into dozens of asymmetric shapes. The key difference from the competitor H.265 (HEVC) lies in motion compensation: instead of using a single motion vector, AV1 applies affine transformation, modeling rotation and scaling of objects between frames, which is critically important for dynamic scenes. Intra prediction uses an extended set of 56 directions, and a Constrained Directional Enhancement Filter (CDEF) is employed for image filtering, suppressing ringing artifacts without excessive blurring. The final stage is entropy coding using multi-symbol prefix codes, which replaced the outdated context-adaptive binary arithmetic coder from H.264, achieving a significant increase in throughput and optimization for multi-threaded processing for real-time operation.
AV1 functionality
- Frame partitioning into superblocks. AV1 introduces superblocks of a fixed 128×128 size, which serve as the root of the recursive partitioning tree. Unlike fixed macroblocks in older codecs, each superblock is processed autonomously, adaptively splitting into blocks ranging from 128×128 to 4×4 depending on the complexity of the local texture.
- Flexible multi-view prediction. The functionality supports up to eight reference frames, including forward and backward projections. Compound prediction allows blending two different predictors in a single block using averaging, weighted interpolation, or wedge masking operations, effectively encoding object overlays and smooth transitions without stitching artifacts.
- Affine global motion. Unlike a translational vector, panning and zooming are described by a single affine model for the entire frame. The encoder transmits transformation parameters (rotation, scale, shift), allowing motion extrapolation onto a static background, radically reducing bitrate for scenes with shaky cameras or slow zooms.
- Spatially varying motion vectors. For non-rigid object deformations, AV1 implements the OBMC (Overlapped Block Motion Compensation) mode. Vectors from neighboring blocks are combined with smooth weight masks at the edges of the current block, eliminating blockiness artifacts on objects with complex kinematics at low bit costs for vector transmission.
- Adaptive optical distortion compensation. Instead of correcting illumination for the entire frame, AV1 applies per-pixel multiplication using a parametric illumination model within the block. This can be a linear luminance gradient or chromatic shift, brilliantly compensating for vignetting and flicker without fragmenting macroblocks into smaller textural residuals.
- Multiple masked prediction. The compound wedge function uses a library of 16 pre-generated non-rectangular masks (wedges) for smooth blending of two reference blocks. The encoder selects a wedge based on the edge geometry of the moving object, forming a sharp boundary without jaggedness at sub-pixel precision within the blending space.
- Directional intra prediction. Intra-frame coding is expanded to 56 angular predictors. In addition to basic directions, AV1 supports directional filtering with a smooth luminance gradient transition, allowing accurate texture extrapolation even for sharp diagonals without staircase effects and residual energy redundancy.
- Recursive intra block copy. The Intra Block Copy mode applies offset vectors to already decoded areas within the current frame. This is a classic approach for screen content, where AV1 achieves compression of repeating interface elements and fonts without transforming them into the frequency domain, preserving glyph sharpness.
- Cross-component color modeling. The CfL (Chroma from Luma) predictor restores chrominance planes by passing the luma signal through linear regression. The encoder transmits only two scale coefficients, forcing chroma to follow the structural contours of luma, eliminating the redundancy of transmitting identical high-frequency patterns in color channels.
- Hybrid primary transforms. Instead of a fixed-precision discrete cosine transform, AV1 uses a set of four transform kernels (
DCT,ADST,FlipADST,IDTX) with independent type selection horizontally and vertically. For rectangular blocks, the kernel library reaches 16-bit precision, minimizing rounding noise and approximating residual distribution. - Wavelet-like identity. The Identity Transform (
IDTX) mode skips frequency decorrelation, transmitting spatial residuals directly. This mechanism is critical for encoding noise-like regions and fine details, where transformation into the frequency spectrum only smears signal energy, creating ringing on sharp transitions. - Low-frequency non-separable transform. LFNIST applies matrix multiplication to the low-frequency region of primary coefficients. Through a trained non-separable matrix of small dimensionality (
4x4or8x8), it collapses the energy of the residual signal, modeling directional textures significantly more effectively than separable methods without a substantial increase in computational complexity. - Multi-symbol arithmetic coder. The AV1 entropy engine operates with an alphabet of not two, but many symbols, up to 16 elements. Together with CDF models based on 15-bit cumulative probability and adaptive statistics updates, this allows accounting for inter-symbol correlations without binarization, increasing speed and packing density.
- Dual ringing artifact filtering. The CDEF (Constrained Directional Enhancement) post-filter searches for the direction of dominant edges and applies a non-linear sigmoidal filter with amplitude limiting. This blurs ringing around contours but does not affect flat areas, preserving the subjective sharpness of the frame after aggressive quantization.
- Segmented loop restoration. The Loop Restoration tool is based on separable symmetric Wiener filters and self-guided projections (SGR). The encoder divides the frame into tiles with a restoration method selected for each, restoring texture destroyed by quantization noise and pulling the frequency response closer to the original.
- Kinetic reference update. The reference frame system is managed through an array of buffer slots. AV1 allows assigning an arbitrary slot identifier to any frame, shifting temporal predictors using refresh and overlay flags, building complex prediction hierarchies for scaling temporal layers without error leakage.
- Adaptive tiling and segmentation. The frame can be sliced into logical rectangular tiles with a single segment header. Raster scan decoding order of tiles is allowed in parallel, with entropy coder probabilities being reset at boundaries. This ensures deterministic fragmentation of CPU core load without loss of compatibility.
- Weighted motion estimation with decay. When predicting weighted alternating frames, a temporal filter algorithm with a noise model is applied. The encoder searches for motion with sub-pixel accuracy and averages blocks, limiting the amplitude of the difference signal through the sensor noise level, effectively suppressing photon noise without accumulating missed vectors.
- Film grain synthesis. Instead of costly encoding of stochastic noise, the encoder sends parameters of its autoregressive model. The decoder generates pseudo-random noise with specified spectral characteristics and overlays it onto the restored clip before display, providing an authentic cinematic film aesthetic at zero bit cost for noise entropy.
Comparisons
- AV1 vs H.265/HEVC. AV1 provides on average 30% more efficient compression at identical visual quality, which is critical for 4K HDR streaming. However, this advantage is achieved at the cost of an exponential increase in encoding computational complexity, making real-time software AV1 processing significantly more resource-intensive without hardware accelerators.
- AV1 vs VP9. Being a direct successor to VP9 technologies, AV1 implements advanced segmentation tools and improved intra prediction, eliminating artifacts on complex textures. The transition from VP9 to AV1 yields a twofold reduction in bitrate, but requires a complete refresh of the decoding device fleet due to the radically different codec architecture.
- AV1 vs H.264/AVC. The AV1 codec demonstrates overwhelming superiority over the outdated H.264, achieving up to 50% bitrate savings through the use of arbitrary-shaped blocks and a directional smoothing filter. For the video conferencing industry, migrating to AV1 means quality HD video even in conditions of unstable communication channels with low bandwidth.
- AVC (Video compression with motion prediction)
- AV1 vs VVC/H.266. Despite VVC’s claims to efficiency leadership, AV1 wins due to the absence of royalty payments, resolving legal barriers to adoption. Technically, VVC shows better results at ultra-high resolutions, however AV1 maintains parity in the bitrate range typical for mass streaming services.
- AV1 vs AVIF (intra-frame application). In static mode (AVIF), AV1 competes with HEIC and JPEG, offering color depth up to 12 bits and HDR support at a smaller file size. Unlike inter-frame mode, intra-frame AV1 compression does not impose prohibitive encoding delays, making the format a practical replacement for outdated image distribution standards on the web.
OS and driver support
AV1 hardware decoding is implemented through specialized fixed-function blocks in the GPU, accessed via graphics drivers providing an API interface (DXVA, Vulkan Video, VA-API, VDPAU); in Windows operating systems, the AV1 profile is built into DirectX 12 starting from version 1903, Linux requires a kernel build with media subsystem support and installation of updated Mesa or proprietary NVIDIA driver, and macOS uses Videotoolbox on Apple Silicon chips with a hardware decoder.
Security
The attack surface is concentrated in the bitstream syntax parser, which processes frame headers, segments, and transform blocks, where an attacker can craft incorrect values for the restoration of symbol elements (CDF), tile sizes, or context modeling parameters, causing out-of-bounds reads; protection mechanisms include strict validation of the arithmetic decoder by limiting the number of extracted bits, forced clipping of frame pointers within the bounds of the reference buffer, and sandbox isolation of software decoders (dav1d) in separate processes.
Logging
The logging procedure at the level of the reference decoder libaom or the high-performance dav1d is carried out through conditional compilation macro definitions, which record OBU (Open Bitstream Units) parsing events, the state of syntax elements at the moment an entropy decoding recovery error is detected, and reference frame integrity flags, outputting structured strings to the standard error stream or file descriptor without stopping decoding on soft errors.
Limitations
The specification sets a maximum resolution of 65536×65536 pixels, a maximum superblock size of 128×128, and a hierarchy of seven temporal layers, while the actual Main, High, and Professional profiles artificially narrow the bit depth to 8, 10, or 12 bits and restrict chroma subsampling to a value of 4:4:4 exclusively for the Professional profile, and the end-to-end encoding latency increases significantly due to the mandatory analysis of partitioning trees up to four levels deep.
History and development
The format was developed by the Alliance for Open Media consortium from 2015 to 2018 as a successor to VP9, borrowing experimental tools from VP10, Daala (Laplacian pyramid coding), and Thor (affine motion), with the first final specification 1.0.0 published in June 2018, and subsequent versions 2.0 and 3.0 added support for scalable temporal coding, improved contour filtering (CDEF) with extended direction search, and parallel processing modes for coarse-grained frame structure.