We measured the three scenes nobody here had ever run, and one of our own facts was wrong
Image to video, reference to video and video to video shipped on the strength of a vendor page. We put a request through each. All three render, and the note saying H3 has no video-to-video endpoint turned out to be false.
Our upstream probe covered text to video and stopped there. Image to video, reference to video and video to video were shipped on the strength of a vendor page — and two of them do not even go to the same upstream. Nothing had ever put a real request through either.
So we did.
All three render
All three came back with a clip. That matters most for video to video, because our own internal fact sheet said H3 has no video-to-video endpoint. The gateway returned a 5.17-second clip in 151 seconds, so that line was wrong and the page is not selling something that cannot be made.
Last-frame-only image to video works too, which settles what the BETA label on that page is allowed to promise.
The fixtures are ours on purpose
The first run failed all three, on inputs borrowed from elsewhere. Borrowed fixtures measure the fixture. Everything above was run on media generated here, at the shapes this site actually submits, through the same request builder the site uses — a probe that writes its own request body is measuring the probe.
Why this is a note and not a release
Nothing changed for a reader today. What changed is that three pages went from "the vendor says so" to "we ran it", and one sentence we had been treating as a fact is now retracted.
That is worth a dated entry precisely because it is not a feature. A page that asks you to check its numbers has to publish the times its own numbers were wrong, or the invitation is decorative.