跪拜 Guibai
← All articles
Frontend · Backend · Developer

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

By 前端之虎陈随易 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The ONNX GPU backend turns FFmpeg's AI filters from a curiosity into a production tool for restoration and upscaling pipelines. Animated WebP support removes a long-standing friction point for anyone building image-heavy web services or content systems.

Summary

The 9.0 release, codenamed "Lei," arrives four months after 8.1 and pushes computation further onto GPUs. A new v360_vulkan filter accelerates 360-degree video transforms, transpose_cuda handles rotation on Nvidia cards, and AMD gains HDR tone mapping, hardware frame-rate conversion, and memory-mapped transfers. On macOS, ProRes RAW decoding moves to VideoToolbox for faster, lower-power transcodes.

Format gaps close across the board. Animated WebP decoding and muxing are now built in, eliminating external tools for GIF-replacement workflows. HE-AAC 960 decoding opens direct ingest of DAB+ broadcast audio. HDR pipelines get SMPTE 2094-50 dynamic metadata passthrough, a Dolby Vision layer splitter, and LCEVC track writing in MP4, completing an encode-decode loop that began in the previous release.

AI video filters cross the threshold from demo to daily use. The ONNX Runtime DNN backend can now execute on a GPU, making super-resolution, denoising, and old-video restoration fast enough for batch work. A niche addition rounds out the release: a dedicated encoder for the Playdate handheld console. Two removals—CELT decoding and pre-11.1 NVENC SDK options—trim legacy maintenance overhead.

Takeaways
Animated WebP decoding and muxing are now native, letting FFmpeg replace external tools for GIF-to-WebP conversion, frame extraction, and compression.
HE-AAC 960 decoding adds direct support for the DAB+ digital radio audio format used in European broadcasting.
A v360_vulkan filter offloads 360-degree video perspective transforms to the GPU, speeding up 4K and 8K VR workflows.
A transpose_cuda filter performs video rotation and flipping on Nvidia GPUs without CPU-GPU data round-trips.
AMD hardware acceleration expands with HDR color conversion, frame-rate conversion, and memory-mapped transfers that reduce copy overhead.
ProRes RAW decoding on Apple platforms now uses VideoToolbox hardware acceleration for faster, more efficient transcodes.
SMPTE 2094-50 dynamic HDR metadata is preserved through the processing pipeline instead of being stripped.
A built-in Dolby Vision layer splitter separates base and enhancement layers without external tools.
The MP4 muxer can write LCEVC enhancement tracks, completing an encode-decode loop for low-bandwidth high-quality video.
The ONNX Runtime DNN backend gained GPU execution, making AI super-resolution and denoising filters fast enough for batch use.
A Playdate video encoder outputs the handheld console's native format directly from FFmpeg.
CELT decoding and pre-11.1 NVENC SDK options were removed, reducing maintenance burden.
Conclusions

GPU offload is now the organizing principle of FFmpeg releases, with Vulkan as the preferred cross-platform path and vendor-specific filters filling the gaps.

Animated WebP support arriving in 2026 shows how slowly foundational format gaps close in ubiquitous infrastructure—WebP animation has been a production format for years.

The ONNX GPU backend is the sleeper feature: AI video enhancement inside FFmpeg shifts from a tech demo to something you can actually run at scale without a separate Python stack.

Adding a Playdate encoder signals that FFmpeg's maintainers treat niche hardware as a first-class target, not an afterthought, which keeps the project relevant across the full spectrum of video-producing devices.

Concepts & terms
LCEVC
Low Complexity Enhancement Video Coding. A codec that layers a low-resolution base stream with an enhancement stream, improving quality at low bitrates without replacing existing codecs.
SMPTE 2094-50
A dynamic HDR metadata standard that carries per-scene brightness and color information, allowing displays to adapt tone mapping frame by frame rather than applying a single static curve.
ONNX Runtime DNN backend
FFmpeg's interface for running neural-network filters (super-resolution, denoising) through the ONNX inference engine. The 9.0 release adds GPU execution support, previously CPU-only.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗