Burn in captions after you generate

A generate is a picture, and sometimes a soundtrack. It is not a captioned file. Transcribe the speech, pick a BASIC or DYNAMIC look, burn the type into the frames, then spot-check names and numbers. Do that after picture lock. Do not ask the video model to typeset the line.

Muted autoplay is the reason this is a finishing pass, not a prompt clause. The feed will often never hear native audio. Viewers read. If the words are not in the pixels, the clip is silent in the only sense that matters.

Generate is not a captioned file

Lock the take first. Captioning a file you are about to regenerate is how you pay twice for the same sentence.

A video row — Veo 3.1, Seedance, Kling Turbo, anything else — returns motion. Some rows also return baked sound. None of them return a timed caption track you can restyle. A prompt that says large yellow captions: "Ships in 12 days" is asking the model to paint letters as texture. You will get cousin spelling, a second language, or a title card over the mouth. You cannot nudge that. You cannot fix a late cue. You cannot swap BASIC for DYNAMIC without regenerating the shot.

Treat the generate as A-roll or B-roll. Treat captions as a later compositor.

Checklist before you open captions:

The take is the one you will ship, not a draft you might throw away.
Duration is locked. A later trim on burned type means recaptioning the whole file.
Ratio is locked. Captioning 16:9 and then cropping to 9:16 puts the line in the wrong place.
Audio is the audio you want transcribed. If you will replace the line with TTS, caption after TTS, not after the generate's leftover mumble.
Safe-area chrome is in mind: TikTok and Reels buttons, YouTube Shorts AI overlay. Placement is a later check, but you should already know the clip is vertical if it is going to a vertical feed.

If any of those five is still moving, stop. Captions are pixels once burned. Burned-in captions cannot be toggled off by the viewer, which is the point on a muted feed, and also why you do not burn a draft.

Native audio does not skip this page. A Veo clip that already talks still needs type on screen for the people who never unmute. The soundtrack and the caption are different objects. One is heard. One is read. Generate the first. Burn the second.

Transcribe, then edit the transcript

Open add captions to video or the AI caption generator on the locked file. The engine behind Versely captions (add_veed_captions) transcribes speech in 165 language codes and times the lines. That is the start of the job, not the end.

ASR is a first pass. It will miss the words you care about most: SKU names, founder names, ingredient lists, prices, model numbers, anything jargon. Read the transcript as copy, not as a log.

Procedure:

Transcribe in the language that is actually spoken. A Spanish take captioned as English is a different film.
Play the file with the transcript next to it. Pause on every proper noun.
Fix spellings to the brand kit: the bottle is Lumina-8, not lumina eight. The founder is Siobhán, not Shivon.
Fix numbers: 12 days, not 12 day; $49, not forty nine if the on-screen claim is a numeral.
Cut filler you do not want on screen. Spoken "um" and "like" do not have to become type. Captions can be slightly tighter than the audio. They cannot contradict it.
Split lines so a muted viewer can read them in the time they are on. Two short lines beat one paragraph.
Do not add a marketing sentence the speaker never said. That is a text overlay, a different tool, and it will desync from the mouth.

If the generate is silent — Kling 3 Turbo is silent on this catalog at 12 credits — there is nothing to transcribe. Add the voice first, or author overlay lines you actually wrote. Transcribing silence produces an empty track and a false sense of being done.

If the generate has native audio and the line is wrong, do not caption the wrong line. Replace the audio, then transcribe. Captioning a baked mistake teaches the viewer the mistake.

Save the edited transcript. You will reuse it if you also need an SRT sidecar for long-form. This page is the burn-in path. The sidecar is a second export from the same words.

Pick BASIC or DYNAMIC, then burn

Style is a billing choice and a brand choice. On Versely captions those are the same switch.

21 BASIC presets (simple, plain, beans, corpo, and neighbours) bill at 1×. Legible, mostly still. This is the default for explainers, accessibility-first cuts, and anything long enough that a 2× look would double the caption line for no reason.
9 DYNAMIC presets (glass, whisper, glide, fusion, terminal, handwritten, backdrop, and the numbered cousins) bill at 2×. Motion in the type. This is the default for a hook you want the muted feed to feel, not merely read.

There is no third tier. Preview a look on a short slice before you render the full file. Ask for preview_caption_style if you are in the agent; do not discover the 2× look after a thirty-second render.

Caption styles is a different catalogue: 45 CSS presets in 9 groups of 5, used by Caption Studio. Do not confuse those 45 with the 21+9 VEED looks. Pick one family for the channel and stay there. The look is as recognisable as the voice.

Placement:

Keep the line out of the bottom chrome and the side button rail. That is the safe area.
Do not cover the mouth on a talking close-up. DYNAMIC looks that travel the frame still need a home that is not a face.
One position per clip unless you have a reason to jump. Karaoke highlighting is a style, not an excuse to park type on the product label.

Then burn. Burning composites the styled, timed words into the frames. The output is an MP4 the feed cannot switch off. That is the short-form object.

Do not burn, then recrop, then wonder why the line sits in the middle of the frame. Crop, then burn. Do not burn, then extend the clip. Extend, then burn. Captions are last among picture jobs, first among delivery jobs.

Spot-check names, then ship

Watch the burned file twice: once with sound, once muted. The muted pass is the real one.

Spot-check list:

Names. Every person, brand, and SKU. If ASR invented a cousin, the burn copied it.
Numbers. Prices, counts, dates, versions. A "3.1" that became "31" is a different claim.
Timing. A line that arrives after the mouth has moved on is a miss. A line that hangs into the next beat is a miss.
Overlap. Type vs face, type vs pack label, type vs platform chrome.
Language. Mixed-language takes often flip one clause. Read it.
Contradiction. If the voice says twelve and the caption says two, the caption wins on mute and the brand loses.

If a name is wrong, edit the transcript and burn again. Do not "fix it in the TikTok caption field." The on-file line is what the muted viewer sees. The post caption is a different box.

If the only error is placement, restyle or reposition and burn again. If the error is the take — a jump, a wrong product, a line you will replace — go back to generate or TTS. Captions cannot rescue a file that is not the file.

When the muted pass is clean, export. Short-form wants this burned MP4. A website player or a broadcast spec may also want a sidecar; that is a second deliverable from the same transcript, not a reason to skip burn-in on the social cut.

If you need a studio seat to run transcription, pick the look, and burn, sign up. The generate you already have is the input. Captions are the pass that makes it readable.

FAQ

Can I prompt captions into the generate instead?

No. A video model paints letters as texture. You cannot restyle them, time them, or correct a misspelled SKU without regenerating the shot. Transcribe and burn after picture lock.

When do I pick DYNAMIC over BASIC?

When the muted hook needs motion in the type and you accept 2× caption billing. BASIC at 1× is the default for long, spoken, or accessibility-first cuts. Preview before you render the full file.

Does a native-audio clip still need burn-in?

Yes, if it will sit in a muted feed. Native audio is heard. Burned captions are read. They are not substitutes. Caption after the soundtrack is the one you want transcribed.