Captions on motion-control takes

Caption the motion-control output once. A driving clip is the performance: pose, timing, sometimes a mouth already talking. The generate is a new take wearing that motion. If you burn captions on the driver and on the output, you double-caption, or you time type to a file you will never upload. Caption the ship file. Leave the driver alone.

The driver is not A-roll unless you meant it to be

Motion control takes two inputs: a clip that carries the movement, and an image of whoever should be doing it. Kling Video V3 Pro Motion Control is the catalog row that names the job: still plus driving clip, 17 credits. Length, beat, gesture path come from the clip. Wardrobe, face, set, and product identity come from the still and the prompt. You are casting a performance, not describing one.

The driver is often a phone clip of a real person dancing, turning, or holding a pack. That clip may have speech, music, or a lyric already in it. It may even have burned-in captions from a previous week. None of that is the ship soundtrack unless you put the driver in the cut. Usually you do not. The output is a new body wearing the pose. Facial performance often does not ride along. Lipsync is a later pass if the mouth has to match a track.

So the audio objects are easy to mix up:

Driver audio. Whatever was on the performance clip. May include speech you will never ship.
Output audio. Whatever the motion-control row actually returned. Do not assume native dialogue. This row is a still plus a driver, not Veo.
Later VO / lipsync. The line you actually want transcribed.

Caption the audio that will be in the MP4 you upload. Transcribing the driver because “it already talks” is a transcript of a person who will not be on screen. Burning that transcript onto the generate is type that does not match the new mouth, the new mix, or the still’s face.

If the driver already had burned captions in the pixels, do not use that file as the driver if you can avoid it. The model will treat letters as texture on a torso. Use a clean performance. Caption after.

Do not double-caption

Double-caption happens in three shapes:

Driver burn + output burn. You captioned last week’s dance, then used that MP4 as the driver, then captioned the generate. The output may inherit chewed type from the driver and a new track. Two transcripts, one frame.
Driver speech + new VO. You lay a product VO on the generate while leftover driver speech is still in the mix. ASR hears both. Karaoke stamps a chorus. The mute viewer reads a lyric on a SKU.
Platform overlay + your burn. TikTok auto-captions on top of a burn you timed to a VO that is not what TikTok heard. Compact block, out of chrome. You cannot turn their overlay off for them.

Fix for (1): clean driver, one burn on the output. Fix for (2): strip or replace the mix before you transcribe. Fix for (3): keep two lines, not a paragraph.

Music-only drivers are a silent-plate problem. There is nothing to transcribe until you add a VO. Running add_veed_captions on a beat is a spend on empty. If the organic hook is a spoken line you will add later, add it, then caption. If the hook is authored type (“the turn is the product”), that is overlay, not karaoke. Snap needs word timestamps on speech.

Lipsync after motion control is common: the body has the dance, the face still needs a line. Caption after the mouth pass, not after the motion row. The mouth rewrites timing. A burn on the pre-lipsync motion take will lead or trail.

Place for a moving pack, not for the driver’s chin

The still may be a product, a mascot, or a person. Motion control will move it. Captions that were safe on frame one will cross a label on frame twenty.

Keep-outs that move:

The pack shot, if the still is a SKU. Hands and orbits put the wordmark through the lower third.
A face, if the still is a character. The driver can dip a chin into the caption band even when frame one was clear.
Platform chrome, which does not move. Bottom ~200px on 1080×1920 is still furniture. Shorts’ May 2026 AI overlay still sits on photoreal generates.

Preview the whole motion, not the first second. A three-second style sample that starts on a still pose will miss the turn that covers the logo. Watch the generate through. Pause on the worst frame. Move type, do not crop the label.

BASIC 1× two-line is the default on product motion. DYNAMIC 2× karaoke wants clean speech the output often does not have. If you add a dry VO after, karaoke can return. If you only have a bed from the driver, keep a static block or skip captions and author overlays.

Safe area is still a snapshot. Recheck in the app. Motion control does not know TikTok’s username.

One ship file, one burn

Order:

Pick a clean driving clip. No burned type in the pixels if you can help it. Rights cleared.
Match the still to the driver (full body if the motion is full body).
Run motion control. 17 credits on the Pro row as published.
Decide the soundtrack: strip driver audio, add VO, or lipsync. Lock that mix.
Picture-lock the output. Trim now.
Transcribe this file. 165 language codes. Read names and numbers to the PDP.
Burn once. BASIC unless a spoken hook on this mix still reads. Preview through the motion, not on a demo card.
Mute phone pass. If type covers the SKU on the turn, move it. If two transcripts are on screen, you double-captioned — go back to the mix.

The AI caption generator and add captions to a video should see the output. The AI video generator is the motion row. Different doors. 21 BASIC, 9 DYNAMIC, 45 Studio CSS presets — none of them should be applied to the driver “to save a step.”

Do not expect to edit the driver after the generate and have captions follow. The output is a new take wearing that motion, not a layer you can retime independently. Recaption if you recut.

Caption the motion-control output once. The driver is the pose. The burn is the ship file.

Burn the output, not the driver, in Versely.

FAQ

Should I caption the driving clip so the generate inherits the type?

No. Inherited type is chewed pixels. Use a clean driver. Caption the output after the mix is locked.

The driver is a talking-head. Will the generate speak the same line?

Do not assume it. Facial performance often does not transfer. If the line matters, lipsync after, then caption that file. Transcribing the driver’s speech onto a silent new face is a mismatch.

Can I keep karaoke from a lyric in the driver’s music?

Only if that lyric is the ship audio and the stamps are clean. Usually the driver music is leftover. Strip it, add the VO you mean, and caption that. Karaoke on leftover lyrics is a second transcript on a product turn.