What just happened in codecs

This release closes parser and demuxer failure modes that could exhaust CPU or memory, loop indefinitely, or overflow the stack; malformed inputs now fail promptly. Improvements across HLS, MPEG-TS, ISOBMFF, Matroska, Ogg, and other formats make this a broad robustness upgrade for applications handling untrusted media.

The release reports a mean decoding speedup of about 1.3×, with gains up to 2×, and cuts memory use to 25–85% of previous levels depending on architecture. It also replaces the old typestate decoder API with an event-based design, making this a performance and integration improvement for Rust JPEG XL users.

Mediabunny 1.61.6 adds workarounds for malformed AVC configuration and shifted video timestamps from Firefox, plus an AAC description issue in WebKit. It also restores server-side encoding performance after a regression, making this a practical cross-browser reliability release for WebCodecs applications.

Across a selected set of images, AVIF generally produced smaller lossy files for photos, screenshots, illustrations, and transparency, while JPEG XL was the clear winner for lossless compression and can shrink existing JPEGs without further loss. The results are specific to the encoders and images tested on one M4 Pro, and the author notes that individual images can be outliers.

Firefox 158 turns on JPEG XL support by default, moving Mozilla’s earlier intent-to-ship into a stable release. With Safari and Chrome also decoding JXL, sites can serve the format to a much broader browser audience while retaining a fallback for older clients.

Mediabunny 1.61.4 fixes the MPEG-TS demuxer reporting incorrect packet timestamps for some malformed streams. Correct timestamps matter to downstream media processing, so this is a focused but useful reliability fix.

VideoMetricsLab 2.0 adds Vulkan support for VMAF v1 across GPU vendors, along with large speedups to other metrics and a command-line interface. Broader GPU coverage makes objective video-quality comparisons less dependent on a particular vendor’s hardware.

Chrome is shipping native JPEG XL image decoding, a substantial step toward making the format practical to serve on the web. Broader browser support gives sites another modern image option without requiring a decoder polyfill.

av1sfm turns the motion vectors already present in AV1 into dense correspondences for COLMAP, avoiding feature detection for every frame. On KITTI sequence 00, its tested Intel GPU path produced 3–6× as many 3D points as SIFT, though the less precise matches make geometric verification a bottleneck: the approach trades precision for density rather than promising a universal speedup.

jxl-rs 0.7.4 fixes a rare abort on possibly valid files and avoids potentially slow table-of-contents decoding on small files. These reliability and small-file performance improvements make the Rust JPEG XL decoder safer to use in applications.

This paper adapts a real-time neural codec to send independently decodable, importance-ranked packets over unreliable channels, preserving task-relevant visual content more gracefully as packets are lost. It is early research, not a drop-in codec, but explores a useful direction for low-latency video links.

Gianni Rosato examines whether JPEG XL is the right choice for web images, a useful counterpoint to the recent momentum behind browser support. For teams choosing a delivery format, decoder availability is only one part of the trade-off.

jxl-rs v0.7.1 reduces binary size by about 100 kB by moving static data behind LazyLock into a Box. For JPEG XL tooling, smaller binaries make integration and distribution easier, especially where footprint and startup matter.

Mediabunny v1.55.7 ships a targeted fix for sample iterators not terminating when exceptions were thrown. That kind of iterator edge case can quietly stall long-running media jobs, so these patch releases are worth taking quickly.

Mediabunny v1.55.5 is out as a patch release. Keeping close to these point updates is practical if you depend on browser media pipelines, because they are where regression fixes and compatibility polish usually land first.