Hacker News Reader

Comment by sovok_x | Hacker News Reader

I'm not okay with this because its author cherry picks images to compare. Those might as well be an outliers. It'd be more convincing when comparing 10k images gallery with a variety of image types, like photos, paintings, sketches, manga, real objects, textures etc and then averaging.

Also subjectively the AVIF lossy compressed flag pic looks better to me because it has less noise, I don't really care about it being more "true" to anything in such use-cases.

View on HN

>I'm not okay with this because its author cherry picks images to compare.

> I don't really care about it being more "true" to anything in such use-cases.

I guess I finally understand the reason why AOM has so many support from a "technical" side of things.

It would also be better to compute the min/max/avg difference of luma/chroma over the whole image than only do a manual comparison by eye.

The whole point of lossy compression is to lose numerical information in a way that doesn't lose visual information. Computing the amplitude of the numerical loss is not very relevant, since it's the visual loss that matters (and those two things are only weakly correlated).

Beware the streetlight effect: https://www.discovermagazine.com/the-sciences/why-scientific...

Such metrics are easy and tempting but one should make sure that they actually correspond to something that matters.

A similar point is made by Zoltan Dienes in his 2008 book “Understanding Psychology as a Science”: https://pbs.twimg.com/media/Ff_BUhTXEAAZZrA?format=png&name=...

Reply on news.ycombinator.com