Captions after Movie mode, not in the prompt

Movie mode plans scenes, generates them, and stitches a film. Captions transcribe speech on a finished file and burn the words into pixels. Mixing those jobs — putting “large yellow captions” in a scene prompt — is how you get cousin spelling in the picture and no timed track you can restyle. Run AI movie maker to picture lock. Then add captions. Not in the prompt.

The agent job is create a multi-scene AI movie: describe the story, review a scene-by-scene draft (each scene can be text-to-video, image-to-video, or first-last-frame), dispatch after you approve, auto-combine. Credits are spent on scenes after approval, not on the plan. None of that typesets a subtitle style.

Movie mode is scenes, not a caption engine

A Movie mode file is a chained object: planned scenes, last-frame handoffs, voice, music, one export. That is a production shape. Burned-in captions are a finishing pass on that object. They are timed to speech, styled from a preset, composited onto frames. They are not a prompt clause.

What Movie mode is for:

• A 30–60+ second brand or founder film that has to hold a character across shots.
• A landing-page hero with more than one beat.
• A short you will cut down later, once the chain exists.

What it is not for:

• Lettering a flavour name on a pack inside scene two. That is a stills typesetter, then image-to-video.
• A hook pack of six openings. That is create_hook_pack.
• Karaoke-style word highlights on a scene that has not been generated yet. Word-timed captions need speech that exists.

Each scene is priced like its own video generate. Prompting type into every scene is how you pay for misspelled pixels you cannot edit as copy. Regenerating scene three to fix a caption is a scene regenerate, not a caption restyle. BASIC caption presets bill at 1×; DYNAMIC at 2× — for transcription work on a file you already have. That is cheaper than a scene, and it is the right object.

The combine step is the other reason to wait. Captions on scene files before stitch will not share a style clock, a safe area, or a single transcript. Burn once on the combined MP4 unless you have a reason to caption a cutdown separately.

Why prompting type into a shot fails

Video models treat letters as texture. A scene prompt that says captions: "We shipped in 12 days" will often return 12 dys, a second language, or a title card that covers the face. You cannot restyle that. You cannot fix timing. You cannot preview caption styles without regenerating the shot.

Silent-autoplay feeds need words on screen. That is still not a reason to bake them into the generate. Most short-form views arrive muted; captions for silent autoplay is that problem. The fix is burned-in captions on the finished cut, plus an authored overlay if the claim is not spoken. Overlay is copy you wrote. Auto captions are speech you transcribed. Scene prompts are neither.

Hook lines you will A/B do not belong in the movie pixels. If the first three seconds will become six variants, generate the film as the body, then run a hook pack against it — or overlay the tested line. Do not ask scene one to typeset six headlines.

Legal lines, URLs, and “not financial advice” are overlay or a lower-third you control. A movie model will invent a cousin disclaimer.

Native-audio scenes make this worse, not better. When the model writes a mix, the spoken line is in the file. Transcribe that line. Prompting a second set of words on screen that do not match the mix is how a mute viewer and an unmuted viewer get two scripts.

Caption the combined file

Order:

1. Approve the scene plan. Regen the bad scene, not the film.
2. Combine. Picture lock the export. If you are still swapping VO or extending a shot, you do not have a parent.
3. Set the language to the language spoken in that file. Add captions transcribes 165 language codes. It does not translate. If the film is dubbed, caption after the dub. That is a different page.
4. Pick a preset. 21 BASIC looks at 1×. 9 DYNAMIC looks at 2×. Preview a few seconds with preview_caption_style before you burn a three-minute DYNAMIC pass. What captions cost across a month of clips is that multiplier at cadence.
5. Read the transcript. Product names, surnames, numbers come back phonetic. On a mute feed the misspelling is the whole message.
6. Check caption timing and safe area at 9:16 on a phone. The bottom strip is chrome. A line on the true bottom is invisible in the feed.

The AI caption generator is the same transcribe-and-burn job with a tool URL. Use it when you are not in the agent. Do not caption the VOD of a movie you still intend to recut. Trim, then caption. A full-length DYNAMIC render you will throw away is the expensive preview.

If a scene has no speech — a silent product insert inside the film — auto captions will not invent a headline. Overlay the claim, or leave the insert silent. Karaoke captions on a silent plate do nothing useful.

Silent autoplay still needs the burn

A Movie mode film that only works with sound on will lose the mute majority on Shorts, Reels, and TikTok. Burn the spoken line. If the hook is not spoken, write it as overlay. Two layers is normal: transcript of the VO, plus one authored claim.

Do not skip captions because the film is “cinematic.” Cinematic is a grade. Mute is a distribution fact.

Do not skip captions because Movie mode’s tool page mentions captions in an export bullet. Finishing still has a style, a language code, and a transcript to proof. Treat that as a pass you run, not a flag you hope was on.

Article 50 labelling is a third layer when the film is realistic synthetic media of people and you publish to the EU. The caption track is not the AI label. Put the disclosure where the template says. Article 50 disclosure is the copy.

When the combined file exists, caption it. Sign up and run Movie mode as scenes, then burn words as captions, not as a prompt.

FAQ

Can I put the caption style in the scene prompt so I skip a pass?

No. Scene prompts are a bad typesetter. You will bake cousin spelling into pixels you cannot restyle. Generate the film, then transcribe and burn a preset you can preview.

Should I caption each scene before combine?

Only if you will ship the scenes as separate cuts. For one film, burn once on the combined file so timing, style, and transcript are one object. Caption cutdowns after you trim them.

The movie has music but no VO. Do I still auto-caption?

Auto captions transcribe speech. A score is not speech. If you need words on a silent or music-led film, write overlay. Do not expect a transcript of a beat drop.