Generation metadata: the phrase that fills nothing
Generation metadata is any note in a file saying it was machine-produced. As a phrase it is too loose to fill a cell here, because it names nothing a reader could look up or check. As of 2026-09-22.
| The term | Generation metadata |
|---|---|
| What it names | Any note in a file stating that it was machine-produced |
| What it is not | A named scheme, a named field, or a described signal |
| Where the register uses it | As the example of a gesture that leaves a cell empty |
Inclusion rule. Words this site uses in a narrow sense, where the ordinary sense would lead a reader to misread a cell. No vendor statement appears on this page. Order. Fixed order: what the word names, what it excludes, then where it is used here.
1A phrase that sounds like a specification
Generation metadata reads like a technical term and behaves like a gesture. It covers a signed credential manifest, a proprietary key in a container, a line in an export log and a claim a vendor has not implemented. Nothing about the phrase distinguishes those, which is why it cannot fill a column.
One entry in this register turns on exactly that. Its guide tells publishers to keep whatever content credentials or generation metadata the tool writes and warns that re-encoding can strip that layer. Both halves presuppose a layer; neither names one, so the cell records not documented and the sentence is kept as a fact.
2What the register demands instead
Three kinds of statement fill the file-contents column. A scheme named so it can be looked up. A metadata key named so it can be checked in a downloaded file. Or a signal whose behaviour the vendor describes. Each hands a reader something to do next, which is the test.
The bar is high on purpose. Several pricing pages here mention metadata somewhere, in the way pages mention things, and if that filled a cell the column would stop separating a vendor that publishes a specification from one that publishes a phrase. The column exists to make that separation.
3The warning attached to the phrase is worth more than the phrase
Re-encoding can strip that layer is a true statement about any attached record, and it is the only place in this register where a page reads as advice to publishers rather than a claim about a product. It is recorded as a fact for that reason, separately from the empty cell beside it.
Read together with the one vendor that publishes the limit of its own named mechanism, the two describe the same risk from opposite ends: a guide warning about an unnamed layer, and a card admitting a named one can be removed.
Nothing on this page is a vendor statement; the values it helps read are on the support table, with the page and the date each one was read from. See also the free tier column, metadata against watermark. Nearby terms: platform label, double label, detection signal.
4From the production side
Next to the marking question in “Generation metadata”, SceneMixer's pages describe its output like this:
SceneMixer writes dialogue into the script analysis, assigns each line to a shot with a speaker and a short delivery note, and the video model performs it — picture and audio generated together, so the mouth forms that sentence in the language you chose.
SceneMixer, AI Dialogue Generator (read 2026-10-13)
Dubbing replaces the audio of a finished clip; the picture was generated for a different line, so mouth shapes, pauses and shot length all belong to the original language. Native-language generation gives the video model the line in the target language before it renders the shot, so the performance, the pacing and the cut points are built for that language.
SceneMixer, Native-Language Dialogue vs. Dubbing in AI Video (2026): Why the Cast Should Speak the Line (read 2026-10-13)