A text overlay is not a caption track

A text overlay is copy you authored and burned. A caption track is a speech transcript, timed. Add a text overlay to a video (add_video_captions) burns one line you typed — hook, badge, CTA — for the whole clip. It does not listen. Sending a mute pack demo to auto-captions will not typeset “WAIT FOR IT” unless someone said those words. Type the line. Place it.

This is overlay versus transcript. It is not a lower third is not a caption track, which is name-and-title cards fighting speech for the same 9:16 band. Different identity problem. Same furniture.

Overlay is copy you wrote

Nobody has to have said it. The tool places a string. You are responsible for the string being true.

add_video_captions (the overlay door, despite the name) takes one piece of text and a required position: top, center, or bottom. Font size, colour, optional background block. list_caption_fonts if the typeface is named. Render burns pixels into the frames. Cost is a standard editor operation. estimate_cost if you want the number.

Use it for:

A hook the speaker never says: “You’re doing this wrong.”
A price, a mg amount, a SKU code that must stay exact.
A badge: “Recipe,” “New,” “Part 1.”
A CTA: “Link in bio,” a URL.
A quote card over B-roll.

Do not use it for:

What the person said. That is add captions to a video: ASR, 165 language codes, 21 BASIC at 1×, 9 DYNAMIC at 2×.
Several different lines at different times. That is timed overlays, a list of start_sec / end_sec.
A name and title for two seconds. That is a lower third — same overlay family, short hold, different layout fight. See that page for the band.
Letters a model painted into a still. That is a generate of type, not an overlay you can restyle. Ideogram can letter a stamp on a pack; you cannot nudge it at 3s.

The pitfall the glossary names: sending a mute clip to auto-captions because you wanted a headline. ASR transcribes speech. Overlay is copy you wrote.

One persistent line, not a transcript

This tool is one string for the whole duration. It does not change at 3s. If you need a question, then a stat, then a punchline, stop and use timed overlays. If you need karaoke of a VO, stop and use captions.

Position is a band, not a free pixel. Centre covers a pack shot. Bottom on 9:16 hits safe area: username, caption field, progress. Top is the usual hook band. Right-rail buttons still eat a badge you parked low-right even if position is “bottom.”

Contrast: white on a white pour fails. Use a background block. Preview on the real file, not a black slate.

If you also run ASR captions, you now have two objects in the frame. They must not share a band. Overlay at top, speech at mid-lower, off the rail. Two lines of speech plus a hook plus a CTA is a poster. Drop one.

Burned-in captions remain the mute-autoplay object for speech. Overlay is the mute object for claims nobody said. A UGC ad often needs both. It never needs both in the same strip.

Word-highlight DYNAMIC looks apply to captions, not to this overlay. Do not expect karaoke on a badge.

YouTube Shorts photoreal AI overlay sits on the picture (YouTube Blog, 27 May 2026). Your hook at top-right can collide with that chip. Long-form labels sit below the player — overlay still lives in the 16:9 frame, usually a lower third that must clear the control bar. Convert ratio before you burn the overlay. Pixels do not reflow.

Distinct from a lower third and from timed beats

Three authored-type jobs, three clocks:

Job	Clock	Tool
Overlay	Whole clip, one string	`add_video_captions` as overlay
Timed overlays	Many strings, you set in/out	`add_timestamped_captions`
Lower third	Short hold, name/title	Overlay family, 2–5s, then off — layout vs captions is the other page
Speech captions	Mouth-timed transcript	`add_veed_captions`

Do not file a CTA as a caption preset. Do not file a transcript as a badge. Do not file a name super as “also captions.”

Title cards that cover the SKU are a crop you burned. Park the line off the pack. That collision has its own how-to. This page is identity: overlay ≠ caption track.

Prompted type in T2V is a fourth, bad door. Veo 3.1 will paint letters as texture. You cannot restyle, retime, or spell-check them without regenerating. Overlay after picture lock.

Legal lines: if a disclaimer must be read, hold it. A whole-clip overlay of a paragraph is unreadable on a phone. Shorten, or accept that the legal lives in the description too. Overlay is not a terms-of-service scroller.

Place after crop, off chrome

Order:

Picture lock.
Ratio. 9:16 / 1:1 / 16:9.
Overlay the authored line on that canvas.
ASR captions if speech needs a transcript, other band.
Export.

Localise the overlay by rewriting the string. Do not ASR a translated VO and expect your English CTA to update. Two jobs. A French hook is often longer; if it no longer fits the band, shorten the copy, do not shrink the type until it is unreadable.

The AI caption generator is a tool surface for speech-shaped captions. This editing task is the overlay. If the agent offers to “caption” a silent clip, specify overlay and the exact line.

Article 50 (2 August 2026) still applies to photoreal picture. A true overlay does not make the clip not-AI. NY GBL § 396-b (9 June 2026) is synthetic performers in ads — not a text badge on a bottle.

QC muted, phone size:

The line is true of the whole clip (because it holds the whole clip). If it is only true for three seconds, you wanted timed overlays.
It misses the pack and the chrome.
It is short enough to finish.
It is not a second caption track.

Burn the authored line in Versely as an overlay. Do not send a mute file to ASR and hope it invents the hook.

FAQ

Is this the same as adding subtitles?

No. Overlay is text you supply. It does not transcribe. Subtitles / captions are speech, timed.

Can I add more than one line that changes over time?

Not with this tool. One fixed line for the whole clip. Changing lines are timed overlays.

Where can the text sit?

Top, center, or bottom — required. Centre on a pack covers the SKU. Bottom on 9:16 hits platform chrome. Pick for the picture you have.