AVIF is an image format derived from I-frames of AV1 video (similar to HEIC/HEIF from H265/HEVC). See also my 2014 Test of BPG , which is an H265 I-frame image format.

Here are some links I've found on AVIF :

AVIF image format supported by Cloudflare Image Resizing
GitHub - AOMediaCodeclibavif libavif - Library for encoding and decoding .avif files
GitHub - googlebrunsli Practical JPEG Repacker
Releases · kornelskicavif-rs · GitHub
GitHub - link-ucavif avif encoder, using libaom directly.
GitHub - xiphrav1e The fastest and safest AV1 encoder.
AVIF for Next-Generation Image Coding by Netflix Technology Blog Netflix TechBlog
Submissions from xiph.org Hacker News
Image formats for the web HEIC and AVIF – The Publishing Project
Comparing AVIF vs WebP file sizes at the same DSSIM
AV1 Codec
AVIF images color losschange AV1

Unfortunately I have not found a standard encoder with recommended settings. There appears to be zero guidance on settings anywhere. Because of that I am using the simplest encoder I could find, Kornelski's "cavif" which has a simple --quality parameter. I run thusly :

cavif --quality=X -o out.avif in.png
avifdef out.avif dec.png
imdiff in.png dec.png 
CALL FOR HELP : if you know better programs/settings for encoding AVIF, please let me know ; in avifenc , what is the min max Q parameter? adaptive quantization appears to be off by default, don't we want that on?

I measure results using my imdiff

I will compare AVIF to what I call "JPEG++" which is JPEG with the packjpg/Lepton back end, and a deblocking decoder (my "jpegdec"). This a rough stand-in for what I think a real JPEG++ should be (it should really have an R-D front-end and chroma-from-luma as well; that's all very easy and unequivocably good).

With no further ado, some results :

(imdiff "fit" is a quality in 0-10 , higher is better)

porsche640.bmp :

PDI_1200 :

Results are a bit disappointing. AVIF is much beter on RMSE but slightly worse on my other two scores. Overall that means it's most likely better overall, but it's not a huge margin.

(I'm sure AVIF is a big win on graphic/text images where JPEG does poorly)

AVIF results here look worse than what I saw from BPG (HEIC). Perhaps better encoders/settings will fix that.

Looking at the results visually, AVIF preserves sharp lines much better, but is completely throwing away detail in some places. There are some places where I see AVIF actually *change* the image, whereas JPEG is always just making the same image but slightly worse.

7z of encoded images to compare (2 MB)

NOTE : the fact that AVIF wins strongly in RGB RMSE but not in my other perceptual metrics indicates that is is not optimizing for those metrics. Perhaps in other perceptual metrics it would show a strong win. The metrics I use here from imdiff were chosen because I found them to be the best fit to human quality scores. Lots of the standard scores that people use (like naive SSIM) I have found to be utter junk, with no better correlation to human quality than RMSE. MS-SSIM-IW is the best variant of SSIM I know, but I haven't tested some of the newer metrics that have come out in the last few years.


Some JPEG Tools

A couple of tips and tools for working with JPEG files. I use :


Of course you can use exiftool and jpegtran but these are rather simpler.

1. Strip all personal data before uploading images to the web.

JPEG EXIF headers contain things like date shot and location. If you don't want Google to scrape that and use it to track your movements and send drones to follow you carrying advertisements, you might want to strip all that private data before you upload images to the net. I use :

jhead -purejpg %*
jhead -mkexif %*
jhead -dsft %*
with the very handy jhead tool. This removes all non-image headers, then makes a blank exif, then sets the exif time from the mod time. The last two steps are optional, if you want to preserve the shot time (assuming the mod time of the file was equal to the exif time). Note if the times are not equal you can use "jhead -ft" to set modtime from exif time before this.

Also note that if you use tools to modify images (resizing, changing exposure, what have you), they may or may not carry through the old exif data; whether that's good or bad is up to you, but it's very inconsistent. They also will probably set the modtime to now, whereas I prefer to keep the modtime = shot time so that I can find files by date more easily.

2. Remove orientation tags

iPhones are the only device I know of that consistently make use of the JPEG orientation tags rather than actually rotate the pixels. Unfortunately lots of loaders don't support these tags right, so if you load the image in a non-compliant loader it will appear sideways or upside down.

(this is a classic problem with data formats that have too many features; inevitabily many implementers only support a subset of the features which they deem to be the necessary ones; then some bone-head says "oh look the format has this weird feature, let's use it", and they output data which most loaders won't load right (then they blame the loaders). This is a mistake of the people defining the format; don't include unnecessary features, and make the practical subset a well-defined subset, not something that forms ad-hoc from use.)

To fix this, you want to read the orientation tag, rotate the pixels, then blank out the orientation tag (you have to do the last step or compliant loaders will doubly rotate). I used to have scripts to do all that with jpegtran, but it's super easy to do with jhead, you just :

jhead -autorot %*

3. Lossless crop

Avoid loading a JPEG, modifying it, and saving it out again. It destroys the image by re-applying the lossy quantization. I think by far the most common modification people do is cropping and there's no damn reason to re-encode for a crop. You can crop at 8x8 granularity losslessly, and you should. (all the web image uploaders that give you a crop tool need to do this, please, stop sucking so bad).

jpegcrop is a GUI tool that provides a decent lossless crop.

FreeVimager is okay too. Lossless crop is hidden in "JPEG Advanced".

We've got JPEG-XL coming soon, which looks great, but all the tech in the world won't help if people like Instagram & Youtube keep re-encoding uploads, and re-encoding every time you modify.

old rants