FFmpeg 9.0 "Lei" Lands With GPU AI Filters, Animated WebP, and a Playdate Encoder

The players, editing software, live streaming tools, and even browsers we use daily likely all run the same thing under the hood: FFmpeg.

It's the Swiss Army knife of the audio-video world, almost everywhere.

On August 4, 2026, it released version 9.0, codenamed "Lei", just about 4 months after the previous version 8.1.

I flipped through the official changelog; there are 17 changes in total.

Let's pick out the key points and talk about them: what scenario each update is used in, and what changes it brings.

Animated WebP can finally be processed

Previously, FFmpeg could only handle static WebP images. This time, both decoding and muxing for animated WebP have been added.

The application scenarios for WebP animations are broad: web animations, stickers, product display images.

At the same quality, its file size is usually much smaller than GIF. Many websites have long been using it to replace GIF.

Previously, converting, compressing, or extracting frames from WebP animations required other tools. Now, FFmpeg can do it with a single command.

People building image hosting services, content management systems, or batch asset processing scripts will benefit directly.

Broadcast audio can also be decoded

HE-AAC 960 decoding has been added. It is the audio format used by European digital radio DAB+.

Previously, FFmpeg couldn't decode this variant, so toolchains for radio recording and rebroadcasting had to take detours.

Now it can be accessed directly. People working with software-defined radio or broadcast-related development will find it useful.

Hardware acceleration continues to ramp up

This part is the highlight of 9.0, with the direction being to let the graphics card do more work.

A new v360_vulkan filter has been added, using the graphics card to handle perspective conversion for 360-degree panoramic videos.

Previously, this filter only had a CPU version, which was very slow for processing 4K and 8K panoramic videos.

For creators working on VR content and panoramic videos, processing speed will now be much faster.

A new transpose_cuda filter has been added, using Nvidia cards for video rotation and flipping.

In scenarios like batch converting vertically shot phone videos to landscape, data no longer needs to shuttle back and forth between the CPU and GPU, making transcoding faster.

AMD graphics cards got three things in one go: HDR color conversion enhancement, hardware frame rate conversion, and hardware memory mapping.

The first two make it more convenient for AMD card users to transcode HDR videos and handle processing like converting 24fps to 60fps.

Memory mapping reduces data copying between the CPU and GPU, improving overall processing efficiency.

On the Apple platform, ProRes RAW video can now use VideoToolbox hardware decoding.

ProRes RAW is a format commonly used by professional cameras. Mac users transcoding this type of footage will find it faster and more power-efficient.

It's clear that FFmpeg is continuing to offload computation to the graphics card, with Vulkan as a key route.

The practical meaning for us is: for the same video processing task, shorter waiting times and more power-efficient machines.

The HDR video pipeline is more complete

SMPTE 2094-50 metadata is now supported. This is a dynamic HDR metadata standard.

Dynamic metadata records the brightness and color information for each segment of a video, which is key to the HDR effect.

Previously, when processing this type of video with FFmpeg, this information was easily lost in the pipeline.

Now it can be preserved and passed through. People doing HDR video transcoding and muxing no longer need to worry about image quality information being lost.

A new Dolby Vision multi-layer HEVC splitting tool has been added.

Dolby Vision video has a two-layer structure: a base layer and an enhancement layer. Many playback devices only recognize the base layer.

Previously, splitting required specialized external tools. Now FFmpeg has this capability built-in.

Doing Dolby Vision video conversion and adapting for different playback devices will be much more convenient.

The MP4 muxer now supports writing LCEVC tracks.

LCEVC is a technology that overlays a layer of enhancement information on a regular video stream to improve image quality, often used for transmitting high-definition images over low bandwidth.

The previous version already supported parsing it. This time, muxing is completed, so the pipeline is considered unblocked for developers doing related encoding tests.

An interesting small update

A new Playdate video encoder has been added.

Playdate is a retro handheld game console with a crank.

Independent developers creating content for this console can now use FFmpeg to directly output its dedicated video format.

AI filters can finally use the GPU

FFmpeg has a built-in set of DNN filters that can do AI processing like super-resolution and denoising.

The application scenarios are very practical: restoring old videos, enhancing low-quality footage, and batch improving video clarity.

Previously, these filters mainly relied on the CPU for computation, and the speed was so slow they were only good for demos.

This time, the ONNX Runtime backend has added support for GPU execution. AI video enhancement has moved from usable to practical.

Two removals

CELT decoding support has been removed. This does not affect Opus audio, so daily use is unaffected.

Deprecated options for NVENC have been removed, and SDKs prior to 11.1 are no longer supported.

The benefit of removing old code is lower maintenance costs and a lighter burden.

If you wrote old scripts that call NVENC, remember to check the command parameters after upgrading.

My view

9.0 is not a major overhaul version, but more of a steady step-by-step improvement.

The direction is clear: embrace the graphics card, embrace Vulkan, and clean up historical baggage.

For ordinary users, FFmpeg's progress will eventually be transmitted to the daily experience through players, editing software, and live streaming tools.

Once various software follows up and integrates 9.0, features like animation processing, HDR video, and AI image quality enhancement will all become faster and better.

Afterword

FFmpeg 9.0, codenamed "Lei", 17 updates, with the main themes being hardware acceleration and format completion.

The upgrade risk is low; only friends using old NVENC parameters need to pay attention.

What do you think of this FFmpeg 9.0 update? Feel free to leave a comment.

Thanks for reading. I am a rural programmer, independent developer, industry observer, Chen Suiyi, the Tiger of Frontend.

WeChat public account 陈随易, personal website: https://chensuiyi.me.