Preview a caption style on 3 seconds

Preview a caption style on three seconds of the real clip before you caption the whole file. preview_caption_style trims the first sample_seconds of your video and runs the matching caption path on that fragment only. It exists so you can compare one or two looks cheaply. Captioning a four-minute podcast clip in Glass to “see if we like it” is a full 2× render used as a mood board.

This week’s radio is not which of the 30 looks is prettiest in the abstract. It is: preview on this background, then commit. The AI caption generator will happily burn the entire file in whatever is selected. Bring the cheap loop yourself.

Why three seconds of this file

A picker thumbnail is not your footage. Contrast that works on a black demo card fails on a white kitchen. Animation that looks fine over b-roll strobes a talking mouth. A bottom preset that clears a demo’s fake chrome sits under the real TikTok stack on your 9:16.

preview_caption_style takes video_url, a style, optional sample_seconds, position, colours, font, preset. Required is the video. The description in the tool registry is explicit: offer 1–2 styles, call this, do not auto-render every style. After they like a look, caption the full video with the matching add-captions path.

Three seconds is enough to see:

Whether the first spoken line is readable.
Whether the type fights the face.
Whether the first cue sits in the safe area.
Whether a dynamic preset’s motion is a 2× you still want.

It is not enough to QA a whole transcript. You still read the full transcript before the full burn. Preview is the look. The full job is the words.

If the hook is not in the first three seconds, set sample_seconds to a window that contains the line you care about, or trim a scratch copy so the preview opens on that line. Previewing the pre-roll of a podcast while the actual hook is at 0:40 teaches you nothing about the hook.

Dynamic presets cost 2× on the full file: 9 looks (glass, whisper, glide, fusion, terminal, handwritten, backdrop, and the rest of that tier) versus 21 basic looks at 1×. Preview is how you avoid buying the multiplier to make an aesthetic decision. The month-of-clips meter is caption_render × (1 or 2) × clips. A preview is not a clip in that leftover. A full Glass burn is.

The cheap loop for this week

Picture lock. Do not preview on a cut you will trim; the first three seconds will change.
Pick at most two candidates. One basic default (Plain, Corpo, a Paper look) and one dynamic only if the channel actually uses motion type.
Run preview_caption_style on the real 9:16 (or whatever you will ship). Watch it muted on a phone.
If both fail contrast, change colour or position. Do not jump to a third family yet.
When one passes, full-caption that preset. Do not “just try” the other on the whole file.

Caption styles is nine families of five. That library is not the same list as VEED basic versus dynamic. You can browse families for taste. You still preview the preset the studio will actually burn.

Fonts: if you care about a specific typeface, list fonts first, then preview with that font_id. Previewing a default and then swapping the font on the full render is how the thing you approved is not the thing you shipped.

Position: top when the mouth owns the bottom; bottom when the face owns the top and you have left margin above the chrome; middle rarely, and only after a preview proves it does not bisect the eyes. Preview is where you learn that, not the live ad.

Do not preview ten styles. Ten previews is a small loop that becomes a habit of not deciding. Two is the brief. The tool description says 1–2 for a reason.

What preview will not save

A wrong transcript. Preview uses the same speech path; if the SKU is wrong in the first sentence, fix the words, then preview again if the look still matters.

A file with no speech. Preview of auto-captions on a mute plate is a blank. If nobody said the line, you are on overlay, not this radio.

An unlocked ratio. Preview on 16:9 and ship 9:16 and the safe area you checked is fiction. Preview the file you will upload.

A series default you already trust. If the channel has used the same 1× preset for three months and this clip is the same talking-head setup, you may skip preview. Preview is for new looks, new backgrounds, and any time you are about to flip 1× to 2×.

Agencies: preview on the client’s actual clip in the kickoff, once. Put the approved preset name in the retainer. Do not preview every episode of a weekly show unless the background changed. The cheap loop is for decisions, not for ritual.

Add captions to a video is the full job after the look is named. Rebuilding a whole timeline to change a preset is wasted work. Preview, pick, burn.

The Monday radio

New look or new background → preview 3 seconds, one or two styles.
Same look, same setup → full caption at the house 1× (or the written 2× exception).
Tempted to 2× → preview is mandatory. If the three-second Glass already annoys you, you do not owe the rest of the file at 2×.

Watch the preview muted. Captions are for mute. A look that only works with the VO in your headphones is not a look.

If you need the agent to run the preview loop on your file, sign up and ask for one or two styles on the first three seconds. Do not ask it to caption the whole cut “in a few options.” That is how 2× multiplies by accident.

FAQ

Can I preview every dynamic preset and pick a winner?

You can, and you should not. The tool is a decision aid for 1–2 looks, not a render farm. Two previews beat nine. If you cannot choose between two, you do not have a style problem. You have a brief problem.

Does a 3-second preview match the full burn exactly?

It is the same caption path on a short trim, so the look, font, and position should match. Timing of later cues will not appear. Approve the look from the preview; still read the full transcript before the full burn.

What if the first three seconds are a silent slate?

Point the sample at speech, or trim a scratch so the preview starts on talk. A preview of a silent slate only tells you whether the preset draws on black. It does not tell you whether it survives the face.