wyohknott image formats comparison includes pik

[Jyrki Alakuijala]

Comparison of modern formats available at http://wyohknott.github.io/image-formats-comparison/

If you have an interest in this, keep in mind a few gotchas:

• the originals are in yuv420 so no benefit or possibly harmful to do rgb or yuv444 on them (pik and flif are the only yuv444 codecs there)
• pik is not compressed to 'tiny' levels nor 'lossless' (as these are not yet supported by the codec)
• this version of pik works well in the range of 1.6 bpp, but many images are compressed only to lower bpp, even 'large' files with < 0.5 bpp (next version of pik will work better in the lower bpp range)
• the codecs to compare are likely AV1-2018, pik and 'original' -- possibly daala, mozjpeg, bpg and webp can be interesting references
• artefacts that the codecs deliver are very different and it may take a while to understand what the codecs are doing to the images
• decoding speed is not a topic of the format comparison, but is a practical issue -- and more important than the encoding speed
• fast encoding of pik is possible, although that is more of a proof of concept in the current reference encoder (even we have seen faster than jpeg results in benchmarks)
• single threaded pik decodes 100x faster than av1 SW reference decoder and 4x faster than webp lossy, ~30x? faster than flif
• pik has a psychovisual definition of the quality based on multiples of the 'just noticeable error' -- and this definition was not used to define what 'large', 'small' etc. means, so quality on 'large' can be very different from one image to the next. On some images large is much worse than small on the other images.
• repeat: pik only works above ~1.6 bpp, elsewhere it is possibly not very rewarding to look at it for now

[zubzer0]

Thank you! pik very nice for "visual-lossless" category (losses visible under upsize)
Perhaps this is the first format that is able to overthrow JPEG in this category of compression.
But with the increase in compression, the probability of artifacts appearing at the boundaries of contrast areas is very high compared to other formats.

Although if you take the example of "fruits oranges", the artifacts even on "Large"
Most likely, due to the high probability of artifacts, "PIK" is not suitable for repeated compressions. As in the case of JPEG ;(
old artifacts + new artifacts

[Jyrki Alakuijala]

Thank you! pik very nice for "visual-lossless" category (losses visible under upsize)
Perhaps this is the first format that is able to overthrow JPEG in this category of compression.
But with the increase in compression, the probability of artifacts appearing at the boundaries of contrast areas is very high compared to other formats.

Although if you take the example of "fruits oranges", the artifacts even on "Large"
Most likely, due to the high probability of artifacts, "PIK" is not suitable for repeated compressions. As in the case of JPEG ;(
old artifacts + new artifacts

Yes. I agree fully! We know that this is a missed opportunity in pik and are working very hard to address it -- without making the visually lossless worse or slowing down decoding a lot.

(((Fruit oranges 'large'/pik is compressed at 0.46 bpp, not a big wonder that it has visible artefacts given that we only optimized in the range of 1.6 bpp so far.)))