Machine shops: sparks are motion, tolerances are captions

Sparks are motion. Tolerances are captions. Kling 3 Turbo Text to Video will give you a silent plate of chips, coolant, and a torch. It will not typeset ±0.02 mm. Do not ask Veo 3.1 to invent the number in the scene either.

The shop's commercial claim is a dimension someone will inspect. Text a video model draws inside the frame has to be right every frame. A softened spark is invisible. A 0 that becomes an 8 for six frames is a scrap part on a feed.

Sparks belong on a silent row

Kling 3 Turbo Text to Video is fast text-to-video. Category is text-to-video. audio is false. Durations run 3s through 15s. Max output is 1080p. Twelve credits is the catalog figure, with discounted billing on the row. requires_image is false. You get a silent plate.

That is the right object for process motion: chips, a weld puddle, a bar feeding. Start at 3s, 4s, or 5s unless the beat is actually longer. Aspect is 16:9, 9:16, and 1:1. Pick it. "Cinematic widescreen" in the prompt will not add 21:9.

Audio is off on purpose. Score later: music, off-screen VO, or a talking row if we see a mouth. If the machinist is on camera and speaking the spec, that speech is a different take, not a Kling paragraph.

A real fixture still belongs in a locked still first, then image-to-video. Turbo text-to-video will invent a cousin mill. Sparks as atmosphere are allowed. "This is our five-axis" needs a photograph.

Specs are authored pixels, not generated glyphs

After picture lock, print the number yourself.

If someone said the tolerance on the take, add captions to the video: VEED transcribes the spoken audio, times the lines, and burns styled captions. It listens. It does not invent a dimension the speaker never uttered.

If nobody said it — the usual shop ad — that is an authored overlay: one fixed line, or timed lines if the callouts change. Type the real drawing callout. Position top, center, or bottom. Burn it. The number is a pixel you control.

Do not "put ±0.02 mm on the part in the prompt" and hope Kling or Veo holds the glyph. On-screen text inside a generate is a 200-frame typesetting problem. A tolerance is specific, small, and a wrong digit is a client escalation.

Veo 3.1 is a native-audio cinematic row (4 / 6 / 8 second clips, 4K tier). Keep it for atmosphere if you need that look. Do not keep it as the typesetter for an inspection dimension.

Picture, then print, then score

Order of operations for a shop clip:

1. Real still of the machine or part, or a short Kling plate of process motion with no numbers in the prompt.
2. Approve the take as a part, not as a brochure.
3. Burn the spec as captions (if spoken) or overlay (if authored).
4. Lay music or VO if the plate is silent and we do not see lips.

Reverse that order and you will regenerate sparks to fix a typo. Sparks are cheap to like and expensive to re-roll once type is baked into a model pass that cannot edit a single digit.

FAQ

Can I ask Kling to write the tolerance in the shot if I make the text huge?

You can ask. You should not ship it. Specific numbers are the case where one wrong letter is not a soft failure. Overlay or transcribe. Do not generate the callout.

If the machinist says the spec, do I still overlay it?

No. That is the caption job: speech in, styled subtitles out. Overlay is for a line nobody said. Do not run both as two copies of the same dimension unless you meant to.

Why not Veo, since it has native audio?

Native audio is not a typesetter. Veo will happily speak a number and draw a different one on the part. Keep Veo for the look if you need it. Print the spec after.

Does Turbo stay silent if I put the spec in quotes in the prompt?

Yes. audio is false. Quoted dialogue will not become a stem. You will get a silent clip, possibly with a mouth that mimes. The number still does not belong in the generate.