GPU changes cover several distinct workloads

The versioned changelog lists an APV Vulkan encoder, an NVIDIA optical-flow interpolation filter called fruc_vulkan, and NVDEC CUARRAY acceleration. It also adds D3D12 DirectML zero-copy inference to the ONNX Runtime DNN backend. These are separate processing paths; their presence is not a claim that every conversion becomes faster.

For an integration team, the useful question is whether its actual pipeline can use a supported path. A video interpolation stage, a decoding stage and a neural filter have different roles. We have not benchmarked this release and report no universal speedup.

Audio, images and transport gain new options

FFmpeg’s list includes an improved AAC encoder, xHE-AAC encoding through libfdk-xhe-aac, ASTC image encoding and decoding through libastcenc, and JPEG Multi-Picture Format handling. The release also introduces a libcurl protocol and additional LCEVC bitstream and transport-stream handling.

These additions expand the available building blocks. They do not prove that a distributor enabled every external library. For reproducible deployment, retain the exact executable version and build configuration alongside the application settings rather than relying on an operating-system package name.

Check removals as well as additions

The changelog explicitly removes the asf_o demuxer. That is a concrete reason for teams with older media inputs to include representative files in migration testing. A release containing new formats can still change a workflow that depended on a removed component.

The download page publishes source archives and identifies external providers of compiled packages. It also provides cryptographic verification information. Developers should distinguish the upstream source release from a downstream executable, whose packaging choices and delivery schedule can differ.

Make the upgrade a measurable media test

Our suggested acceptance check is a small collection of real production inputs: a typical video, a problematic legacy file, and audio material used by the application. Compare duration, track selection, synchronization and expected metadata before considering throughput. A job that finishes faster with missing content is not an equivalent result.

Record the operating system, driver and exact command or application configuration with each run. If an optional GPU or codec path is unavailable, report that explicitly instead of treating a software fallback as a test of the new acceleration. This separates confirmed support from assumptions and makes later regressions easier to reproduce.