Our speed numbers were invented. We ran the probe, and the real ones are faster.
Every render-time figure on this site was a stand-in until today — including the ones printed under copy promising they were measured. Twenty-two timed runs later they are real, and two of them were wrong by more than half.
Until today, every render-time figure on this site was made up.
Not estimated from a vendor page — invented, as a placeholder, so the speed sections could ship with a number in them instead of a hole. They then sat under copy saying things like "measured on this site" and "what we measured rather than what the model card claims", which makes them worse than the hole they were filling. The module holding them said so at the top, in a warning nobody outside the codebase could read.
We ran the probe. Five rounds against the engine this site actually defaults to, 22 timed runs, request out to finished file.
The real numbers, and they are faster
| Shape | We were claiming | Actually |
|---|---|---|
| 5s · 480P | 6.4s median | 4.9s (5 runs) |
| 5s · 768P | 8.7s median, 20.9s P95 | 6.9s median, 8.9s P95 (12 runs) |
| 15s · 768P | 22.6s median | 13.4s (5 runs) |
The invented figures were slower than the truth in every row, and the tail was badly wrong: we were advertising a P95 of 20.9 seconds on a shape whose slowest run out of twelve was 10.0.
Sample size is now 22 and says so. It used to say 412, which was a number nobody had ever run.
Two things that are not about speed
The upstream does not enforce the five-second floor. A four-second request —
below what H3 Max accepts — comes back 200 accepted and then simply never
completes. It does not refuse. Every guard on this site that assumed the gateway
validates on our behalf has to be our own.
Image to video does not complete on the Turbo post-train. A first frame, a last frame, or both: all three accepted, none finished inside ten minutes. That is a real limitation of the engine the site defaults to, not a transient, and the image-to-video page has to say so.
What is still made up
Being honest about the fix means being honest about its edges. A probe can time a request; it cannot watch sixty clips and score whether the lips landed on the syllable. Still stand-ins, and now listed by name in the code rather than covered by one flag that reads "measured":
- the render time for the base H3 model — only Turbo was probed
- lip-sync, camera-move and beat-order hit rates
- face-hold across clip lengths
- the verified language list
- the lifetime clip count and rejection share
And one edge on the numbers above: the clock stops when the file is fetchable, one step before our own re-store to storage. They are a floor on the wait, not the whole of it.