Added 13
- Experimental --fast-decode 3 decode-speed level for still images, enabled by the ENABLE_FAST_DECODE_3 CMake option, with decoder-complexity terms in the mode decision RD cost
- Timed film grain tables via --fgs-table and per-picture film grain parameters through a new FILM_GRAIN_EVENT API
- Low-delay tile parallelism
- Opt-in CBR pre-encode frame skipping via max_allowed_consecutive_frames_skips and EB_NoErrorFrameSkipped
- Screen content tools for RTC: IntraBC on I-frames and palette on intra blocks of inter frames
- Signaling of AV1 frame_id for application-managed LTR streams, with LTR anchors kept in a dedicated half of the DPB
- max_hierarchical_levels setting to size buffer pools for runtime mini-GOP size changes
- MOBILE_BUILD CMake option (build.sh --mobile-build) producing one 8-bit library for RTC, VOD and AVIF at a fraction of the default binary size
- CONFIG_ENABLE_* switches to compile out unused coding tools
- Native 8-bit CDEF kernels (AVX2, SSE4.1) and int16 lowbd forward transforms (AVX2, AVX-512) on x86
- SSE4.1 tier for intra predictors, CfL, SAD, Hadamard, palette, and transform kernels on x86
- Neon kernels for fwht4x4, highbd variance, k-means, OBMC weighted prediction, down2_symeven, and interpolate_core on Arm
- Neon and SVE2 Hadamard SATD and psy distortion kernels on Arm
Changed 12
- Improved quality vs. speed tradeoffs across all presets, with about 5-8% improvement in cycle efficiency at similar quality levels and about 10-11% run time reduction in multi-threaded mode
- Sped up MD pruning, RDOQ and coefficient cost, MV cost and subpel search, loop filter parameter derivation, and block copies
- Run anti-aliasing-aware screen content detection only on I-frames in multi-threaded mode
- Lossless coding fixes and speedups in recode, RDOQ, and light PD1
- Re-encoding of key frames is allowed on max frame size conditions
- RTC preset retuning: hierarchical levels 1-2, CDEF strength selection, and ME segmentation right-sizing at 540p and below
- Further binary size reduction for RTC and minimal builds through const-folding, dead-code stripping, and a PGO build script
- Reduced stack usage by hoisting large buffers off the stack
- Optimized quantize, memcpy, block copy, full distortion, and sub-pixel variance kernels on x86
- Optimized VMAF tune, variance, and sub-pixel variance kernels on Arm
- MIDR_EL1 is only read where the kernel advertises it
- Stricter configuration validation for switch frames with all-intra, forward key frames with CBR, and RTC build settings
Fixed 8
- Bitrate inversion at fractional CRF/QP above 63
- Multi-instance and TSan data races, and nondeterministic output with --enable-overlays 1 or --lookahead 0
- A low-delay deadlock, an Apple teardown SIGTRAP, and single-core encoding without --lp
- Out-of-bounds reads, a high bit depth Hadamard overflow, UBSan findings, and NaN/overflow in film grain and temporal filtering noise estimation
- Hidden-frame T.35 metadata output and recon output changing the deblocking level search
- The --tier parameter being ignored when writing the sequence header
- Encoder/decoder mismatch when quantization matrices were enabled on frames coded losslessly
- Visual Studio 2026 and Windows Server 2016 issues
From SVT-AV1
[4.3.0] - 2026-11-01
Encoder (general)
- Improved the quality vs. speed tradeoffs across all presets:
- ~5-8% improvement in cycle efficiency at similar quality levels
- ~10-11% run time reduction in multi-threaded mode
- Speed optimizations in MD pruning, RDOQ and coefficient cost, MV cost and subpel search, loop filter parameter derivation, and block copies
- Run anti-aliasing-aware screen content detection only on I-frames in multi-threaded mode
VOD / Random Access / Still Image
- Added experimental
--fast-decode 3(ENABLE_FAST_DECODE_3CMake option): a decode-speed level for still images, with decoder-complexity terms in the mode decision RD cost - Added timed film grain tables (
--fgs-table) and per-picture film grain parameters through a newFILM_GRAIN_EVENTAPI - Lossless coding fixes and speedups (recode, RDOQ, light PD1)
- Fixed bitrate inversion at fractional CRF/QP above 63
RTC / Low Delay
- Added low-delay tile parallelism
- Added opt-in CBR pre-encode frame skipping (
max_allowed_consecutive_frames_skips,EB_NoErrorFrameSkipped) - Added screen content tools for RTC: IntraBC on I-frames and palette on intra blocks of inter frames
- Signal AV1
frame_idfor application-managed LTR streams, and keep LTR anchors in a dedicated half of the DPB - Added
max_hierarchical_levelsto size buffer pools for runtime mini-GOP size changes - Allow re-encoding of key frames and on max frame size conditions
- RTC preset retuning (hierarchical levels 1-2, CDEF strength selection, ME segmentation right-sizing at 540p and below)
Build
- Added
MOBILE_BUILDCMake option (build.sh --mobile-build): one 8-bit library for RTC, VOD and AVIF at a fraction of the default binary size - Added
CONFIG_ENABLE_*switches to compile out unused coding tools, and further binary size reduction for RTC and minimal builds (const-folding, dead-code stripping, PGO build script) - Reduced stack usage by hoisting large buffers off the stack
x86
- Added native 8-bit CDEF kernels (AVX2, SSE4.1) and int16 lowbd forward transforms (AVX2, AVX-512)
- Added an SSE4.1 tier for intra predictors, CfL, SAD, Hadamard, palette, and transform kernels
- Optimized quantize, memcpy, block copy, full distortion, and sub-pixel variance kernels
Arm
- Added Neon kernels for fwht4x4, highbd variance, k-means, OBMC weighted prediction, down2_symeven, and interpolate_core
- Added Neon and SVE2 Hadamard SATD and psy distortion kernels
- Optimized VMAF tune, variance, and sub-pixel variance kernels
- Only read
MIDR_EL1where the kernel advertises it
Bug fixes and documentation
- Fixed multi-instance and TSan data races, and nondeterministic output with
--enable-overlays 1or--lookahead 0 - Fixed a low-delay deadlock, an Apple teardown SIGTRAP, and single-core encoding without
--lp - Fixed out-of-bounds reads, a high bit depth Hadamard overflow, UBSan findings, and NaN/overflow in film grain and temporal filtering noise estimation
- Fixed hidden-frame T.35 metadata output and recon output changing the deblocking level search
- Fixed the
--tierparameter being ignored when writing the sequence header - Fixed an encoder/decoder mismatch when quantization matrices were enabled on frames coded losslessly
- Stricter configuration validation (switch frames with all-intra, forward key frames with CBR, RTC build settings)
- Fixed Visual Studio 2026 and Windows Server 2016 issues
- General code cleanup, CI and test improvements, and reformatting with clang-format 22