Notices by Gianni Rosato (gianni@disobey.net)

Hey everyone! fmetrics, a fast open-source perceptual metrics toolkit that runs on the CPU, is here: https://github.com/halidecx/fmetrics

fmetrics includes support for IW-SSIM, MS-SSIM, CVVDP (via fcvvdp), SSIMULACRA2, & Butteraugli. All are faster than their reference implementations while maintaining competitive MOS correlation – see the blog post (https://halide.cx/blog/fmetrics) for more info.

All code is Apache 2.0. Enjoy!!

Making Software: Image Compression
https://www.makingsoftware.com/chapters/image-compression

Cool to see WebP addressed here too!

@pirboazo oops, sorry! Let me contact the host

@tommi @yawnbox Hey Tommi, nice to meet you!

I think something like https://github.com/fabianwimberger/immich-convert-originals might be close to what you're looking for? I haven't vetted it (I don’t batch process images for my personal use case), but you can check whether everything is preserved with `exiftool`.

Broadly, there are 3 common image metadata formats you’ll run into: ICC (color information), EXIF (descriptive information about the photo, like your camera specs), and XMP (custom XML). If you procure a test file, run it through whatever tool/script/etc you’re thinking about using (libjxl, libavif, or something else), and check the before & after with `$ exiftool <image>` to ensure the metadata is the same, you should be good to go with any photos similar to the test case (e.g. you know they have the same metadata fields & file format).

Good luck, lmk how it goes!

fcvvdp 0.2.0! https://github.com/halidecx/fcvvdp/releases/tag/0.2.0

Highlights: ~9-12% speed boost for single-threaded, task threading (220% faster than the reference implementation with 8 threads on my 13700k), FFmpeg support via patch. Check it out and lmk what you think!

@barrbrain wow, definitely going to look into this, thanks! Very cool!!

New Halide blog post on post-processing for chroma as it is relevant to compression efficiency: https://halide.cx/blog/chroma-handling

Seems like you can get a roughly 2% efficiency improvement with 4:2:0 decoder output just with smarter post-processing for chroma. Pretty neat!

#avif #webp #jpeg #transcoding #compression

@rachelplusplus Best of luck, I’ve consistently been a big fan of all of your image compression blog posts!

SVT-AV1 4.0.0 is here!
Changelog: https://gitlab.com/AOMediaCodec/SVT-AV1/-/releases/v4.0.0
SVT-AV1-PSY's porting efforts are fully complete, and this release also features huge AVIF improvements. Maybe the most exciting release to date by my standards – SVT-AV1 is finally a real perceptual encoder!

#av1 #avif #compression

Happy Holidays! Introducing fcvvdp, a faster CVVDP implementation: https://github.com/halidecx/fcvvdp

Compared to the reference implementation running on the CPU, fcvvdp is faster (~15%) despite being single-threaded. It sees stronger wins with user time (up to ~15x) and memory usage (>90% less memory).

#compression #av1 #video

I spoke with Julio Barba at VideoLAN Dev Days in London!

Here are our slides, discussing SVT-AV1-PSY: https://giannirosato.com/files/svt-av1-psy_vdd_2025.pdf

I'll find access to the video & post it soon!

#av1 #avif #ffmpeg #vdd

@eeeps thanks for the shout out, glad ur liking my work!!

Introducing oavif: faster target quality image compression

https://giannirosato.com/blog/post/oavif/

oavif can be up to 63% faster than traditional target quality encoders. Learn how in the attached blog post!

#av1 #avif #compression

fssimu2 has its first release! https://github.com/gianni-rosato/fssimu2/releases/tag/0.1.0

fssimu2 is a fast SSIMULACRA2 implementation in Zig, with support for distortion map output. It is ~23% faster than the reference implementation at ~40% less memory usage, while being 99.97% accurate. Give it a look if you're interested!

New Halide blog post!

https://halide.cx/blog/consistency

Quality and speed define an image encoder's compression performance. Consistency is a close third, and easily overlooked in image encoder design. What value does it provide, and how can we measure it?

#avif #webp #jxl #jpegxl #compression

Anyone using rss/atom readers can now type `giannirosato.com` into their reader to get my website's atom feed

I got to interview Julio Barba from Google for the Halide blog: https://halide.cx/blog/julio-barba-interview/

Go check it out if you're interested in what the next decade of multimedia tech might look like!

Completely redesigned my website with Zola (https://www.getzola.org) – check it out at https://giannirosato.com

Let me know if any links are broken if you link to my site – it should be fully backwards compatible with my old link hierarchy

Tune 3: Still Picture is coming to SVT-AV1: https://gitlab.com/AOMediaCodec/SVT-AV1/-/merge_requests/2489

Formerly Tune 4 in SVT-AV1-PSY. Thank you Julio!

Exciting #avif news – libaom’s tune iq (derived from SVT-AV1-PSY’s Tune 4) is used across The Guardian’s website now. They say they serve mostly AVIF, with JPEG XL shipped when supported. Very exciting stuff!