count .png-s in the web. There's a lot. And .jxl's lossless is much better than png's. Basically any image with sharp lines (screenshots, image of a button etc) or gradients should be lossless.

This is outdated thinking. AVIF is really good at sharp lines & gradients. https://hacks.mozilla.org/2026/08/intent-to-ship-jpeg-xl/#:~...

Not at lossless, though.

also, don't forget - actual video format is better for animated stuff. Current default approach is .webm which is standardized .mkv container + vp9 (or av1) video + opus audio (usually no audio to as a feature parity to gif). IMO animated picture formats like .avifs are very backwards and don't add much

One advantage to animated AVIF is that it supports an alpha channel, which AV1 doesn't. But I guess that just points to how short-sighted it was for AV1 to omit alpha channel support.

zstd with better compression level would be nice - these numbers are not really comparable since both time and compression level are too different

Tell that Apple and Android. I have 32 bit native software which can't run on modern devices at all.

I'm sorry to hear that. Could it be ARM related?

Apple dropped 32 bit support in specific OS version with no way to run it anymore sans downgrading the OS - so it's not hardware restriction

It's just to reduce the maintenance burden of backwards compatibility.

> early hours of the morning

which TZ?

> There are no off-peak hours.

Probably US west coast hours, but there's no public stats

2AM CST

Dunno. Crease almost invisible when screen is on and not pitch black ( using Honor V3 while actual generation is V6).

Does that also mean owner must re-run very model ever released by them to detect anything?

1) That's a lot more entropy to gather, means sample size needs to be huge.

2) this also means check must be run against every user's ID in their database, 10^9 more expensive

That makes sense, though it sounds like it would still be feasibly cheap to only track a handful of users or organizations

Do i understand correctly - to check watermark you need full model weights, of all org models. So running the check is basically the same as running every model once? That's really expensive

I don't know the details of their implementation, but in general, if the text is not watermarked, you will probably fail the test after a few tokens. If it is, you will have to run the entire text, or, you can just run it until you have some degree of confidence.

That would be Sv, not SV.