Correction, 2026-10-01: an earlier version said generating silent is half the price; that holds on Veo 3.1 Standard (224 against 448 credits) but silent Kling v3 Standard is about a third cheaper (57 against 86). It also gave caption line lengths and safe-area advice without a source, and said generated frames "consistently" break on-screen text, which we have not measured. The figures now cite Netflix, Meta and the 8frame registry, and the closing line uses our standard wording.
Caption for silent viewing by designing the video to work without sound in the first place, then adding captions as reinforcement rather than as a rescue. Captions pasted onto a video that depends on its voiceover produce a wall of text nobody reads. The videos that hold muted attention are the ones where the picture carries the meaning and the captions confirm it.
Setup
Decide two things before you caption anything: whether captions are burnt in or uploaded, and whether the piece is narration-led or picture-led. Both decisions change how you write them.
Burnt-in captions are part of the video file. They always appear, you control the typography completely, and they cannot be turned off or translated by the platform.
Uploaded caption files (SRT or VTT) are separate, can be toggled, are used for accessibility tooling, and can be auto-translated. They render in the platform's own style, which you do not control.
The practical answer for most social video is both: burnt-in for design control and an uploaded file for accessibility and reach.
The rules that make captions readable
Two lines maximum, no more than 42 characters per line. That is the limit in Netflix's English subtitle style guide; go shorter on narrow vertical frames. Three-line captions cover the picture and get skipped.
One idea per caption card. Split at natural clause boundaries, not at arbitrary character counts, because a caption broken mid-phrase reads as two half-thoughts.
Hold each card at least one second. Cards flashing faster than that cannot be read, which is common in auto-generated word-by-word captioning styles.
Place them in the safe area. Platform UI covers the edges of vertical video, so captions sitting at the very bottom get hidden behind buttons and usernames. Meta asks advertisers to keep the top 14%, the bottom 35% and 6% on each side of a 9:16 Reels ad free of text and logos, so sit captions just above that bottom band rather than at the true bottom.
High contrast, always. A background plate, a stroke, or a shadow. Captions that rely on the footage being dark enough will fail on the one shot that is not.
Do not caption what is visible. If the frame shows the product being opened, caption the point, not the action.
Designing for sound-off before captioning
The caption layer works when the video is already legible muted.
Show the subject in the first frame, not after a title card.
Use text for the hook, not narration. The first two seconds have to communicate without audio because that is when the decision to keep watching happens.
Demonstrate rather than describe. Sound-off video is a visual medium, and anything explained only in voiceover is invisible to anyone watching muted.
Design one key frame that carries the message on its own, for the viewer who watched with the phone face-down.
Where generated video changes things
Two practical notes.
Many video models now generate sound with the picture. On 8frame, Gemini Omni Flash always generates audio, while Veo 3.1 and Kling v3 let you switch it off, which matters for cost: a silent Veo 3.1 Standard 8-second clip is 224 credits against 448 with audio, and silent Kling v3 Standard at 5 seconds is 57 against 86. If the piece is caption-led with a music bed, generating silent halves the price on Veo 3.1 Standard and takes about a third off on Kling v3 Standard.
Second: do not rely on on-screen text generated inside the frame. A model can misspell or garble it, and you cannot fix one letter without regenerating the shot. Generate clean footage and add captions in the edit, where you control the typography and the spelling.
FAQ
Do captions actually improve performance? We do not have a first-party number to cite, and we do not print uplift percentages we cannot trace to a primary source; treat any you see quoted with suspicion unless it names one. What is documented: WCAG 2.2 success criterion 1.2.2 (Level A) requires captions for prerecorded audio in synchronized media, and Meta lists captions as recommended for Reels ads.
Burnt-in or uploaded captions? Both. Burnt-in for design control on social, uploaded for accessibility and platform translation.
How fast should captions move? No faster than one card per second, split at clause boundaries. Word-by-word styles look energetic but give the reader less time with each line; Netflix's English guide allows up to 20 characters per second for adult programs.
Sources
- Netflix, English (USA) Timed Text Style Guide: 42 characters per line, two lines maximum, up to 20 characters per second for adult programs, checked 2026-10-01
- Meta, Instagram Reels ads specs: keep 14% top, 35% bottom and 6% each side free of text and logos; captions recommended, checked 2026-10-01
- W3C, Understanding WCAG 2.2 success criterion 1.2.2, Captions (Prerecorded), Level A, checked 2026-10-01
- 8frame audio options and silent prices (Veo 3.1 Standard 224 vs 448, Kling v3 Standard 57 vs 86, Gemini Omni Flash audio always on): 8frame back tool registry cost functions, origin/main, read 2026-10-01
Design the piece muted, generate silent where the model lets you, and add the caption layer in the edit.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.