qoi2-bikeshed
qoi2-bikeshed copied to clipboard
nigeltao
More benchmark suite image variety (esp. re alpha)
Per #14
The test images are linked to from https://phoboslab.org/files/qoibench/ (with misc/lenna.png removed).
and
x kodak misc screenshots textures wallpaper total average QOI_COLOR (40) 0.000% 9.986% 0.000% 0.000% 0.000% 1.997%
It's not surprising in retrospect that, other than misc, nothing uses 5-byte QOI_COLOR codes, as the source pixels are fully opaque: they're photographs and screenshots.
Should we source some more PNG-images-with-alpha for our test suite? Perhaps some logos or SVG-to-PNG conversions??
Similar but not exactly the same: https://github.com/phoboslab/qoi/issues/20
I think a good set of test images with natural alpha channel data is the 213 icons in The Tango base icon library which is "released to the Public Domain", rendered at 512x512 (scripts in the package uses librsvg)
P.S. Please excuse that I'm a stranger with a new account.
Great suggestion. I have created https://github.com/nigeltao/tango-icon-library-pngs to hold the 512x512 PNGs.
No excuse necessary. Strangers welcome!
Note sure where to post this. This 11kb PNG image encodes to a 256kb QOI. I feel like this is a worst-case scenario for that codec. I didn't generate that image for that purpose, I was just testing PNG vs QOI. So far, pretty underwhelmed with QOI compression size.
- The original PNG is 10790 bytes (11K).
- The official QOI implementation encodes this in 262170 bytes (257K). I used phoboslab/qoi commit c04a975 from 2021-12-24.
- The unofficial qoi-lz4-d28 experimental format encodes this in 1065 bytes (1K).
Some more good images to test against: https://github.com/jhcloos/test_images
@nigeltao it's nice that qoi-lz4 does better, but better entropy-coding doesn't really fix the lack of vertical intra prediction.
And some more images: https://www-abclinuxu-cz.translate.goog/blog/programy/2022/5/jxl-from-tree?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=cs&_x_tr_pto=wapp .