Slower speed bigger size

[avidseeker]

Sample image:

for i in $(seq 0 10); do avifenc -s $i --qalpha 0 -q 40 "$f" -o $i.avif; done

$ du -b *.avif
3920    0.avif
3896    10.avif
3853    1.avif
3867    2.avif
3843    3.avif
3883    4.avif
3878    5.avif
3862    6.avif
3779    7.avif
3746    8.avif
3896    9.avif

I find it surprising that the slowest speed lead to the highest size.

Being a bit familiar with libjxl, I thought this corresponded to cjxl -e effort flag which is explained as follows:

Controls the amount of effort that goes into producing an “optimal” file in terms of quality/size. That is
to say, all other parameters being equal, a higher effort should yield a file that is at least as dense and
possibly denser, and with at least as high and possibly higher quality.

Recognized effort settings, from fastest to slowest, are:

I'm curious about what's happening here. Is there a way to make avifenc smart about the size/speed and immediately calculates to go for speed 8 is optimal?

[maryla-uc]

Hi,
Comparing lossily compressed images is not as simple as looking at the size, because the visual quality of the encoded image is also different. The file produced at speed 0 (the slowest speed) is indeed larger, but it's also slightly higher quality. What you can do is to draw a plot with size on the X axis and some quality metric (the simplest being PSNR) on the Y axis. This allows you to visualize the size-quality trade off.

Here is a simple plot for your example image

aom00 is aom at speed 0. It's toward the top left which is where you want to be.

In general, we have data that does show that all other things being equal, speed 0 produces better files. I.e. they are smaller for the same visual quality, or better quality for the same size. Here is a chart that keeps visual quality constant and shows that s0 (speed 0) is indeed smaller (and slower).

That being said, I agree that this behavior is counter intuitive. I think we should also print some quality stats in avifenc.