Text inside the frame: the one layer nothing can fix later
Text inside the frame is any writing a model draws into the picture: a sign, a note, a phone screen. Subtitles sit on top of it and dubbing does not touch it, so it is the only language layer that cannot be repaired after the shot exists. As of 2026-09-22.
| Layer | Can change it | What would be needed instead |
|---|---|---|
| Subtitle | No | A new render of the shot |
| Dub | No | A new render of the shot |
| Script change before generation | Yes | Nothing further |
| Shooting the text as its own insert | Yes | One short regeneration per market |
Inclusion rule. The ways a production can change writing that appears inside a generated picture. Two of the four require the picture to be made again. Order. Cheapest attempted fix first, which is also the one that fails.
1It is a rendering problem wearing a language costume
Drawing legible letters is about glyph shapes, spacing and joining rather than about grammar or pronunciation. A model can speak a language convincingly and be unable to write it, and the reverse is equally possible.
That is why this register keeps such claims outside all three columns. Counting a frame-text claim as evidence about dialogue would be the same category error as counting a translated menu.
2One claim about it exists in this register
A vendor launch post read for this register claims multilingual text generation inside frames, with accurate placement, naming two languages and then saying others. It is the only statement of its kind here.
Which scripts it covers is not stated, and scripts are where the capability breaks. A model trained largely on Latin and Chinese characters is not evidence about a diacritic-heavy or right-to-left script.
3The cheapest arrangement is to avoid it
Where the story does not require legible writing, keeping it out of the frame removes the whole problem. Where a sign has to be readable, shooting it as its own short insert means each market regenerates a few seconds rather than a scene.
That is production advice this register can support, because it follows from what the layers can and cannot reach rather than from any vendor claim about quality.
A definition, not a measurement: no vendor figure appears here that is not also in the register. The one frame-text claim here is recorded beside the empty voice cell it sits next to.
Nearby terms: Burned-in text, Reading speed. The whole vocabulary is at terms; nothing on this page is a claim about a named product.