you know how JPG gets artifacts from reëncoding over and over again because it's lossy? well I was wondering what the artifacts for lossy JXL looks like. since reshared photos are often either text posts or cats, I grabbed a screenshot from the #bradley tag (first post, second post) and started exporting it as lossy JXL repeatedly at the worst setting, maxError = 15. it turns out that
- a single run already crunches it up pretty bad! the first run is closer to the one on the right than the original
- it took six runs to reach a fixed point—the seventh run was indistinguishable
- to me, at least, text starts looking like it's from poorly-generated image AI...
the background colour changing hue was surprising to me too. why is it a completely different shade?
for those of you with browsers with JXL support here's the authentic™️ JXL'd image
