Add Captions

The Add Captions node puts captions into a video. It transcribes the speech in the video, or takes a transcript you connect, and burns the words into the picture. You can show a clean static subtitle or one of five animated styles that follow the speech word by word. You also control the font, the outline, the position and the number of words on each line.

When to use it

• You want subtitles on a talking-head or narration video.
• You want animated, word-by-word captions for TikTok, Reels or Shorts.
• You want accessible captions on any video with speech.
• You want karaoke-style lyrics over a music video.

Quick start

Inputs

The output, Video, is the video with the captions burned in.

Settings

Fonts

The Font list has 24 fonts:

Rubik, Heebo, Cairo and Tajawal cover Hebrew and Arabic. Hebrew and Arabic captions run right to left. The direction follows the language of most of the captions, so a Latin brand name at the start of a Hebrew clip does not flip its lines.

Caption styles

The five animated styles are called kinetic styles in the app.

Looks

A look bundles several visual choices, so a caption reads well from one setting.

• Controls override single parts of a look. With Outline (TikTok), changing only Spoken word color keeps the font, the capitals and the outline.
• The spoken-word color shows on three styles. Only TikTok Words, Karaoke and Word Highlight mark one word at a time. Subtitle and Word Pop use the look's font, outline and casing in one color.
• Font size depends on the look's font. Montserrat Black in uppercase is about 30% wider per character than Inter in mixed case. So the same Font Size fits fewer words per line under Outline (TikTok). Choose Clean, a narrow font such as Bebas Neue, Anton or Oswald, or a smaller size.
• The outline grows with the text. When the look draws the outline, its width is 10% of the font size, rounded, and at least 2 pixels. Half of it sits outside the letters, so the visible rim is about 5% of the font size: a 6-pixel outline at size 64, a 3-pixel outline at size 32. Set Outline width to choose your own, or 0 for none.

How captions are grouped into lines

Word Highlight, Karaoke, Bouncy and a transcribed Subtitle show one line at a time: not one word at a time, and not the whole transcript as one block.

• A line takes as many words as fit in about 85% of the frame width, at the chosen font size and font.
• A line ends early at the end of a sentence, marked by ., !, ? or …, or at a pause of 0.5 seconds or more between two words.
• A phrase too long for one line is split into balanced lines, so its last word is never left alone. For example, OK SO I BUILT / A WORLD IN / NODARO STUDIO. instead of ending on a lone STUDIO..
• A line stays on screen for up to 1.5 seconds after its last word, and the next line replaces it the moment the next line starts. Pauses inside a line and short pauses between lines never leave a blank frame. A silence longer than 1.5 seconds clears the caption.
• Inside the line, the effect moves from word to word: the highlight, the fill or the bounce. During a pause, the last spoken word stays active. A subtitle line has no per-word effect.

The two styles that do not draw lines follow the same rules:

• Word Pop keeps each word on screen until the next word starts, for at most 1.5 seconds after the word ends.
• TikTok Words never runs a page across a sentence end or a pause of 0.5 seconds or more. Each page stays until the next page starts, for at most 1.5 seconds after its last word.

Max words per line

Max words per line caps the number of words on one line, or on one TikTok Words page. It works on top of the rules above: a line still ends at 85% of the frame width, at a sentence end and at a pause. The cap can only make lines shorter. Use 1 or 2 for short, punchy captions, or leave it empty to fill the width.

For example, with no cap, Word Highlight shows the lines OK SO I BUILT, A WORLD IN and NODARO STUDIO. With Max words per line set to 2, it shows OK SO, I BUILT, A WORLD and so on, each held until the next one starts. A phrase with an odd number of words ends with a one-word line.

The cap counts words, not caption entries. A phrase-level caption with more words than the cap is split into shorter parts, and its time is divided between them by the length of their text.

Animate

Animate switches the per-word motion of the animated styles. Turning it off freezes the movement but keeps the line grouping, the line holding and the spoken-word color.

For a fully static caption with no color change either, also set Spoken word color to the same value as Color.

Position

Every position anchors the caption block by the edge nearest to the frame edge, so a caption that wraps to more lines grows inward and is never cut off. The placement is the same for every style.

Vertical position sets the center, not an edge. At 85%, the block's center is at 85% of the height and its bottom hangs lower. On a 1920-pixel-tall frame, 83.5% puts the center at about 1,603 pixels. A value around 65% sits below a face and above an app's bottom buttons.

Where the words come from

In the editor, Add Captions gets its words in one of two ways:

• It transcribes the video. With nothing connected to Transcript, the node transcribes the video's speech with Incredibly Fast Whisper. Make sure the speech is clear.
• It reads a connected transcript. Connect the transcript output of Transcribe or Apply EDL to the Transcript input. Apply EDL moves every word through the cut, so the captions stay aligned with a re-cut video, to within about 80 milliseconds across a crossfade.

A transcript needs word timings

A connected transcript must carry the timing of each word. A transcript without words is refused before any credits are reserved.

The Whisper engine of Transcribe returns phrases, not word timings. If a Transcribe node on Whisper feeds this node, directly or through Apply EDL, the workflow stops before anything runs or is charged. The message names the Transcribe node. Switch that Transcribe node to an engine with word timings, such as Incredibly Fast Whisper or ElevenLabs STT. The same check covers chains inside a sub-workflow. Only a chain that crosses a sub-workflow boundary is refused at this node, after the transcription has run.

Word-level captions

With a transcript connected and an animated style chosen, Word-level captions decides how the transcript's words become captions:

• On (the default) makes one caption per word, which the per-word styles need.
• Off groups the words into lines first. A line ends at a change of speaker, the end of a sentence, a silence or a maximum number of words. Use it for calmer, line-at-a-time captions.

Max words per line applies either way.

Transcription engines

The engine does not change the price of Add Captions. The animated styles need word timings, so they need Incredibly Fast Whisper or ElevenLabs STT. Subtitle needs only phrase timings, so it works with every engine.

Frame rate of the result

An animated caption render keeps the source video's frame rate. The rate is rounded to a whole number, such as 24 for 23.976 or 30 for 29.97, and kept between 15 and 60 fps. Three cases render at 30 fps instead:

• The source frame rate cannot be read.
• The source has a variable frame rate.
• The clip is longer than the render's frame limit of 108,000 frames. At 60 fps, that is a clip longer than 30 minutes.

A plain static subtitle always keeps the source frame rate.

Factory presets

Add Captions ships seven presets in the Caption Styles folder. Each one sets the style, the position, the font size and the color, and transcribes the video.

See Presets for how to apply, save and share presets.

Credits

Add Captions has two prices:

The settings panel has no text field, so a node set up in the editor always captions the speech or a transcript, and each run costs 50 credits. The 30-credit subtitle is a fixed text sent through the API, MCP, SDK or CLI. The transcription engine and the styling controls do not change the price of an animated style.

Tips

• Match the style to the content. Use Subtitle for professional content and Word Highlight or TikTok Words for social media.
• Mind the platform's buttons. The default Bottom position already clears the TikTok and Reels buttons. For vertical social video, Center also works well.
• Keep text readable. White text over a dark picture reads best. Use colored text on a light picture.
• Correct the words first. For the most accurate captions, transcribe with Transcribe, correct the text, then time it with Forced Alignment.
• Caption last. Put Add Captions after Video Overlay, so logos and cards stay under the captions.

Troubleshooting

The workflow stops before it runs and names a Transcribe node. That Transcribe node uses the Whisper engine, which returns no word timings. Switch it to Incredibly Fast Whisper or ElevenLabs STT.

Lines hold fewer words than expected. The Outline (TikTok) look uses a wide font. Choose Clean, a narrow font such as Bebas Neue, or a smaller Font Size.

A caption is cut off at the frame edge. Use a named Position instead of a Vertical position near the edge. Named positions anchor the edge of the block, so extra lines grow inward.

From the API and MCP

Code and AI assistants can do more than the editor. POST /v1/add-captions, the add_captions MCP tool, the SDK's client.media.addCaptions(...) and the CLI all accept the same controls as the settings panel, plus the following:

• Your own text. On Subtitle, text is burned in as one fixed block for the whole video and is never transcribed over. A line break in the text forces a new line. On an animated style, text is only a fallback: it is used when the transcription finds no words, and is spread evenly over the video.
• Your own word timings. captions is a list of caption entries, each with text, startMs and endMs. For the animated styles, give one entry per word.
• The transcription engine. transcribe_provider chooses incredibly-fast-whisper (the default), elevenlabs-stt or whisper. An animated style with Whisper and no other source of words is refused before any credits are reserved. Set auto_transcribe to false to turn transcription off; a request with no source of words at all is refused.
• Different treatments per time range. segments applies a different style and look to each time range of the same video, in one call. See Per-segment captions.

When a request carries several sources of words, the node uses the first of: captions, a transcript, text on Subtitle, then transcription.

Field names differ by surface. The REST body and the SDK use maxWordsPerLine, the MCP tool uses max_words_per_line, and the CLI uses --max-words-per-line.

Per-segment captions

Each segment has a start and an end in milliseconds, must not overlap another segment, and can set its own style and look controls. A segment without its own words uses the shared captions or the transcription, limited to its time range. Caption times are always measured from the start of the video, not from the start of the segment. Per-segment captions always use the animated render, so they cost 50 credits.

For example, these arguments to the add_captions MCP tool show a large outlined phrase at the top for 3 seconds, then one word at a time at the bottom:

{
  "video_url": "https://example.com/clip.mp4",
  "look": "outline",
  "segments": [
    { "start_ms": 0, "end_ms": 3000, "style": "subtitle", "position": "top", "font_size": 96, "text": "Same face, every shot. No re-prompting." },
    { "start_ms": 3000, "end_ms": 20000, "style": "word-pop", "position": "bottom", "font_size": 48 }
  ]
}

• Words at a boundary. A single word belongs to the segment that contains its start, so it never appears twice. A phrase that crosses a boundary is split there, and each part shows in its own segment's style. A part shorter than 250 milliseconds is dropped, unless that would lose the whole phrase.
• What a segment inherits. Color, background color, position, vertical position, Animate and Max words per line come from the top level when the segment does not set them. A segment that sets its own position does not inherit the top-level vertical position. The look controls, such as font, weight, outline, spoken-word color and uppercase, are inherited only by a segment without its own look. A segment that names a look starts fresh from that preset.