Test new `-tune=ssimulacra2`

[juliobbv-p]

Recently, libaom landed a new SSIMULACRA2 tune, specifically optimized for still pictures: https://aomedia-review.googlesource.com/c/aom/+/194662. It's currently available on the main branch.

The new tune consists of a set of optimized defaults (QMs, deltaq, rdmult and luma/chroma allocation tweaks, etc) that improve SSIMULACRA 2 scores, and in our preliminary testing, it also has a more favorable subjective quality profile.

It'd be nice to have SSIMULACRA 2 results (in addition to the default tune) so we can have a more complete look at how other metrics react (especially with butteraugli and DSSIM).

[y-guyon]

Here are my findings on a dataset from https://www.compression.cc with multiple codecs at default encoding effort. Let me know if you expect significantly different results at other speeds.

Web-quality range and 4:2:0

Restricting AVIF bpp to [0.03-1.03] range and images to 2-3 megapixels to match the 10th to 90th percentiles at https://github.com/webmproject/codec-compare/wiki/Bits-per-pixel-of-Internet-images.

For the same Butteraugli score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 13% bigger than AOM_TUNE_SSIM.

For the same DSSIM score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 6% smaller than AOM_TUNE_SSIM.

High-quality range 70-90 and 4:4:4

Restricting to AVIF quality range 70 to 90 is arbitrary.

For the same Butteraugli score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 17% smaller than AOM_TUNE_SSIM.

For the same DSSIM score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 14% smaller than AOM_TUNE_SSIM.

[juliobbv-p]

Thanks @y-guyon for running the tests! I took a quick look, and I can already tell they're very informative.

Tune ssimulacra2's changes mainly deal with bit redistribution, so I wouldn't expect wildly different results at other speeds.

That said: I'd also recommend testing speed 3, as it enables rectangular partitions, larger transforms, and restoration filters; while still being fast enough for some production scenarios. It'd be interesting to see how the tune's tweaks interact with the larger available tooling repertoire.

Edit: I'd also suggest testing 10-bit speed 6, as the additional internal precision can make a significant difference at preventing banding, which SSIMULACRA 2 heavily penalizes. At very high quality, SSIMULACRA 2 scores can increase by a whole point or more.

[gitoss]

Here are my findings on a dataset from https://www.compression.cc with multiple codecs at default encoding effort. Let me know if you expect significantly different results at other speeds.

Here's another comment: AOMediaCodec/libavif#2412 (comment)

[y-guyon]

That said: I'd also recommend testing speed 3, as it enables rectangular partitions, larger transforms, and restoration filters; while still being fast enough for some production scenarios. It'd be interesting to see how the tune's tweaks interact with the larger available tooling repertoire.

At speed 3 the gains are in the same ballpark:

Web-quality range and 4:2:0

Restricting AVIF bpp to [0.03-1.03] range and images to 2-3 megapixels to match the 10th to 90th percentiles at https://github.com/webmproject/codec-compare/wiki/Bits-per-pixel-of-Internet-images.

For the same Butteraugli score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 14% bigger than AOM_TUNE_SSIM.

For the same DSSIM score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 6% smaller than AOM_TUNE_SSIM.

High-quality range 70-90 and 4:4:4

Restricting to AVIF quality range 70 to 90 is arbitrary.

For the same Butteraugli score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 15% smaller than AOM_TUNE_SSIM.

For the same DSSIM score

Plot

• AOM_TUNE_SSIMULACRA2 leads to files 11% smaller than AOM_TUNE_SSIM.

Edit: I'd also suggest testing 10-bit speed 6, as the additional internal precision can make a significant difference at preventing banding, which SSIMULACRA 2 heavily penalizes. At very high quality, SSIMULACRA 2 scores can increase by a whole point or more.

My dataset is 8-bit. Do you have a 10-bit corpus with a permissive license to recommend?

[juliobbv-p]

Thanks @y-guyon for the follow-up! It's reassuring to see the gains for SSIMULACRA 2 hold up proportionally for s3.

My dataset is 8-bit. Do you have a 10-bit corpus with a permissive license to recommend?

My apologies, I should've clarified. By "10-bit", I was thinking of using the CLIC dataset, but converted to 10-bit before encoding (e.g. by specifying -d 10 in avifenc). Enabling the AV1 HBD codepath (done by the 10-bit conversion) can greatly reduce banding and improve SSIMULACRA2 scores for images with subtle gradients.

For example, this image of a flower against a blurry background from CLIC:

8-bit	10-bit

• 8-bit: size 152 KB, SSIMULACRA 2 score 83.72
• 10-bit: size 138 KB, SSIMULACRA 2 score 85.86

The 10-bit image is 9% smaller, yet it scores 2 points higher. Both images were encoded at speed 6, with tune SSIMULACRA2.

Close visual inspection of the 8-bit image reveals subtle banding/blocking that's not present in the converted 10-bit image:

8-bit	10-bit

Here are the source and encoded files for reference: 8bit vs 10bit.zip