strukturag
Expose SVT-AV1 parameter avif=1 in libheif for significantly lower memory usage when encoding still images
While running benchmarks for still-image AVIF encoding using libheif + SVT-AV1, I noticed that the memory usage is significantly higher than when using SVT-AV1 directly. It appears that libheif does not forward the SVT encoder parameter --avif 1, even when explicitly specified using svt:avif=1.
According to the SVT-AV1 logs, enabling avif=1 activates optimizations for still images (reduced parallel picture sets, different rate-control mode, etc.).
Below is a reproducible example.
1. Encoding via libheif (high RAM usage, SVT warning shown)
Command:
/usr/bin/time -v heif-enc --avif -p "threads=1" -p "speed=9" --quality=44 -o test.avif parrot-jpg.jpg
Relevant log:
Svt[info]: Number of PPCS 42
...
Svt[info]: BRC mode / rate factor : CRF / 35
...
Svt[info]: Only a single picture was passed in. Consider setting avif=1 for improved efficiency and reduced memory usage.
...
... some more /usr/bin/time logs
...
Percent of CPU this job got: 100%
Maximum resident set size (kbytes): 82856
2. Encoding via SVT-AV1 directly (much lower RAM usage)
Equivalent parameters using ffmpeg + SvtAv1EncApp:
bash -c "ffmpeg -y -i parrot-jpg.jpg -pix_fmt yuv420p -f yuv4mpegpipe - | /usr/bin/time -v SvtAv1EncApp -i stdin --preset 9 --crf 35 --lp 1 --avif 1 -b testsvt.avif"
Relevant log:
...
Svt[info]: Number of PPCS 2
Svt[info]: BRC mode / CQP Assignment : CQP / 35
...
... some more /usr/bin/time logs
...
Percent of CPU this job got: 65%
Maximum resident set size (kbytes): 34216
In summary being able to pass --avif 1 to SVT is helping reduce RAM and CPU usage, the resulting image is smaller in size, and the encoding time is also lower. avif=1 is a documented SVT-AV1 parameter intended specifically for still-image encoding.
Given the above benefits it would be very helpful if libheif exposed a way to pass avif=1 to SVT-AV1, e.g.:
-p svt:avif=1
or simply enable it automatically when encoding a single image with --avif.
If this change aligns with libheif’s design goals, I’m happy to prepare a PR.
Additional info
In the above example I'm using the following image but I am able to reproduce RAM and CPU usage differences with any image:
And here is SVT's documentation for avif parameters.
I tried to add it (edcb909c796a9ef8c7c1a572f72cb84e8340f68b), but it seems that avif=true is currently not working:
https://gitlab.com/AOMediaCodec/SVT-AV1/-/issues/2245
I also get the error
Svt[error]: Error: A picture has already been sent. The library only supports one picture in AVIF mode.
even though only one picture is encoded.
Thanks so much for looking into this and sharing the link to SVT issue @farindk Looks like some early work for fixing the issue was recently merged. Hoping the fixes are released soon :)