FFmpeg 9.0 Supercharges Jellyfin and Every Self-Hosted Media Server With Vulkan, Animated WebP, and ONNX Runtime

August 3, 2026. FFmpeg 9.0 is out. The multimedia library that powers Jellyfin, Plex, Emby, and virtually every self-hosted transcoding pipeline ships a major release with 17 new features. The official CHANGELOG lists advances that directly impact your homelab media server’s performance.

If you self-host a media server, this release isn’t cosmetic. Vulkan, ONNX Runtime, and expanded AMD AMF hardware support redraw the map of home transcoding. Here’s what changes and what you need to enable.

Vulkan Steps Out of Transcoding’s Shadows

GPU-accelerated transcoding is the backbone of any self-hosted media server. For years, the choice was binary: VA-API (Intel/AMD on Linux) or NVDEC/NVENC (NVIDIA). Vulkan Video — Khronos’ cross-vendor video decode/encode API — was the perpetual third option: always in development, never production-ready.

FFmpeg 9.0 shifts the landscape with three Vulkan additions:

• The v360_vulkan filter transposes 360° video using the GPU via Vulkan — essential if you serve immersive content from your homelab.
• APV (Android Platform Video) Vulkan hardware acceleration is now supported, directly benefiting Jellyfin on Android TV users or mobile clients connecting to a self-hosted server.
• A brand-new transpose_cuda filter rounds out the NVIDIA side.

This isn’t yet the VA-API drop-in replacement the community has been waiting for, but Vulkan Video becomes a first-class citizen in FFmpeg. For homelabbers running AMD GPUs or Intel’s newer Arc cards, the trajectory is confirmed.

AMD AMF Extends Its Reach

AMD Radeon cards have long been the poor cousins of Linux hardware transcoding compared to NVIDIA. FFmpeg 9.0 dramatically narrows this gap:

• The AMF Color Converter (vf_vpp_amf) gains extended HDR capabilities, enabling proper HDR10/HLG to SDR conversion — the single most common scenario when a non-HDR client plays from a 4K library.
• The AMF Frame Rate Converter (vf_frc_amf) debuts for hardware frame rate conversion, useful for sports broadcasts or 60 fps content.
• AMF hardware memory mapping is now supported, reducing CPU↔GPU copy overhead and therefore transcoding latency.

For a Jellyfin server powered by a Radeon RX 6600 — a popular choice at roughly $200 on the used market — these optimizations translate to lower CPU-per-stream usage and better stability during long transcoding sessions.

ONNX Runtime: AI Upscaling Lands in FFmpeg

This is the most disruptive addition in this release. FFmpeg 9.0 integrates an ONNX Runtime backend with GPU execution provider support. Translation: you can now apply AI models — super-resolution, denoising, colorization — in real time through the FFmpeg pipeline, without external scripts or hacked-together Docker containers.

The canonical homelab scenario: a DVD rip encoded at 480p that you want to upscale on the fly to 1080p for a mobile client. With an ONNX super-resolution model (such as Real-ESRGAN exported to ONNX), FFmpeg can now run the inference directly — and accelerate it via the GPU if the execution provider is configured.

The integration isn’t yet a plug-and-play -vf onnx_upscale=model=esrgan.onnx filter, but the backend is in place. The maintainers of Jellyfin and Tdarr (automated transcoding) will be able to leverage it in the coming months.

Animated WebP, Dolby Vision, and Other Gems

The full feature list includes several other advances relevant to self-hosting:

• Animated WebP decoder and demuxer: no more forced WebP-to-GIF transcoding on your Immich instance. FFmpeg reads them natively.
• HE-AAC 960 (DAB+) decoding: valuable if you record digital radio via TVHeadend.
• Dolby Vision multi-layer bitstream splitter: multi-layer Dolby Vision rips (profile 7, common on UHD Blu-ray) can now be cleanly split — a direct upgrade for MakeMKV → Jellyfin pipelines.
• SMPTE 2094-50 support (dynamic HDR metadata): HDR-to-SDR tone mapping gains precision, improving transcoded stream visual quality.
• LCEVC muxing in the MP4 container: the MPEG-5 LCEVC enhancement codec can now be embedded in standard MP4 files.
• Playdate encoder and muxer: for the three people serving content to the Playdate handheld console.

What to Do in Your Homelab

If you use the official Jellyfin/Plex/Emby Docker images, do nothing. The maintainers will bump FFmpeg in their next releases. Manually replacing the FFmpeg binary inside a container is the fastest way to break your stack.

If you compile FFmpeg yourself (bare-metal hosts, rolling distributions), compiling with --enable-libonnxruntime and --enable-vulkan activates the two headline features. Here are the flags you want:

./configure \
  --enable-libonnxruntime \
  --enable-vulkan \
  --enable-amf \
  --enable-libvpl \
  --enable-libdav1d

If you use Tdarr or Unmanic for automated transcoding, schedule a worker node upgrade as soon as the official containers ship FFmpeg 9.0. The CPU gains from AMD AMF transcoding and the availability of the ONNX backend justify the update cycle.

Verdict

FFmpeg 9.0 is not a disruptive release — it introduces no universal new codec. It’s a consolidation release that strengthens three pillars of self-hosted transcoding: Vulkan (cross-vendor), AMD AMF (brings AMD closer to NVIDIA parity on Linux), and ONNX Runtime (opens the door to integrated AI upscaling).

If you run a self-hosted media server, your action plan is straightforward: wait for your media distribution’s update (Jellyfin, Plex, Emby, Tdarr), then enable the hardware codecs matching your GPU. Your homelab transcoding is about to get faster and look better without a hardware upgrade.

References

• FFmpeg Git, « Changelog — version 9.0 », accessed August 5, 2026.
• Phoronix, « FFmpeg 9.0 Released With More Vulkan Acceleration, Animated WebP & More AMD AMF », August 3, 2026, by Michael Larabel.
• FFmpeg Website, « News — FFmpeg 9.0 Release », accessed August 5, 2026.