Comment by gyan
7 hours ago
Changelog:
ffprobe -codec option
EXIF Metadata Parsing
gfxcapture: Windows.Graphics.Capture based window/monitor capture
hxvs demuxer for HXVS/HXVT IP camera format
MPEG-H 3D Audio decoding via mpeghdec
D3D12 H.264 encoder
drawvg filter via libcairo
ffmpeg CLI tiled HEIF support
D3D12 AV1 encoder
ProRes Vulkan hwaccel
DPX Vulkan hwaccel
Rockchip H.264/HEVC hardware encoder
Add vf_scale_d3d12 filter
JPEG-XS parser
JPEG-XS decoder and encoder through libsvtjpegxs
JPEG-XS raw bitstream muxer and demuxer
IAMF Projection mode Ambisonic Audio Elements muxing and demuxing
Add vf_mestimate_d3d12 filter
xHE-AAC Mps212 decoding support (experimental)
Remove the old HLS protocol handler
Vulkan compute codec optimizations
swscale Vulkan support
LCEVC metadata bitstream filter
Add vf_deinterlace_d3d12 filter
ffprobe: only show refs field in stream section when reading frames
ProRes Vulkan encoder
LCEVC parser
LCEVC enhancement layer exporting in MPEG-TS
TIL: JPEG XS - an image and video codec that offers both visually and mathematically lossless quality for low latency implementations.
Additionally, JPEG XS compressed content is indistinguishable from the original uncompressed content.
https://en.wikipedia.org/wiki/JPEG_XS
I've had great results using JPEG-XS to transport video for colour grading in feature film & TV post production. At 3:1 or 4:1 compression ratio is effectively lossless.
It is patent-encumbered though, you have to pay license fees to deploy it.
Isn't the point of JPEG to have lossy compression for your photos that still looks fine? As opposed to something like PNG, which has lossless compression
"JPEG" is short for Joint Photographic Experts Group, an ISO/ITU group that creates a lot of imaging standards. The JPEG image format you're thinking of is only one of the formats they've created.
The Joint Photographic Experts Group manages many standards, generally each called "JPEG [something]". The one we most commonly call "JPEG" is just one of them.
Reading that it looks like the point of JPEG-XS is to have near-lossless compression for raw photo and video data while having extremely high throughput.
1 reply →
> Additionally, JPEG XS compressed content is indistinguishable from the original uncompressed content.
It can be indistinguishable, as long as you stick with lossless or very low compression ratios. It falls apart at typical JPEG XL compression ratios.
Not royalty free, unfortunately.
We use JXS when latency is critical. Most h24/265 decodes will have a 10 frame glass-glass delay, JXS drops that to 3 or 4, at a cost of bandwidth (our UHD jxs streams are 1.5gbit rather than 200mbit for hevc)
Yeah, we've been deploying JPEG-XS for high bitrate streaming for a while.
A lot of our customers are moving their grading systems into data centres and streaming the images over IP back to their grading suites.
I've got it down to less than 1 frame for encode-transport-decode, but you've still got to copy the image to an SDI card and wait for that to clock out.
> gfxcapture: Windows.Graphics.Capture based window/monitor capture
> This source provides low overhead capture of application windows or entire monitors. The filter outputs hardware frames in d3d11 format; use hwdownload,format= if system memory frames are required.
This would strongly alter my plans if I were to develop an OSS Discord alternative. Chromium originally looked like a better core to start with largely due to its mature screen capture API. WebRTC is the other big thing, but there are other ways to do that. Native desktop apps (i.e., not browser based) are beginning to look much more compelling to me now.
If you were doing this, consider cribbing from https://github.com/obsproject/obs-studio/tree/master/plugins... which offers a variety of solutions including some rather exciting looking process injection (called "game" there).
I wonder if "entire chat app functions as OBS plugin" would work? Would solve the AV streaming side of the functionality.
You could always use Windows.Graphics.Capture directly.
What I want to know is how much of these were written and/or debugged using AI tools and which ones? Using which workflow?
For that's an actual project, with countless uses, on countless machines.
Show me the AI. I want to see what AI has generated in those.
(btw I pay religiously my Claude Code subscription plan)
Wait, are you https://www.gyan.dev/ffmpeg/builds/ ?