Subtitles and decoration
Subtitles communicate spoken words; titles, labels and decorative overlays serve different purposes. Plan each layer so that it does not obscure the image or imply speech that never happened. Russian terms: субтитры, оформление.
Create an episode’s design
Section titled “Create an episode’s design”Use Decoration (Оформление) after the video is assembled. New nodes start with V2 · по смыслу проекта and produce the finished video on the server. One video connection is enough. V2 carries the connected scenes and their edit mappings, views actual images over the movie’s timeline, and uses speech recognition and sound accents to decide how graphics can strengthen its meaning.
There is no required style selection. Without a connected design system, the agent creates an original reusable style. Optional script, SEO and generator recommendations help it understand the audience and goal. Scene intentions describe what was planned; they are checked against actual images and speech. Image sampling can miss short events between samples, and long videos have sparse visual coverage. The result’s Why these overlays section explains the decisions. Full 30–60 minute production workflows are not yet a verified capability.
The older explicit starting directions remain available for projects already using them:
| Direction | Starting point |
|---|---|
| Editorial | Open space, contrasting serif and sans-serif type, restrained titles and margin notes. |
| Signal | Clear headings and precise labels that call attention to objects or evidence. |
| Kinetic | Large type, emphatic words and short movements, with pauses that leave the picture visible. |
Start with one required connection; add only the material the episode needs:
| Connect to Decoration | Supply |
|---|---|
| видео | The finished video to decorate. |
| контекст (optional) | The script, product facts or a generator’s recommendations for titles, labels and moments to emphasize. |
| VTT (optional) | Reviewed subtitle lines and their times, usually from Subtitles. |
| дизайн-система (optional) | A Design system shared by the series. |
Use пожелания к оформлению for the purpose of the graphics: what the viewer should understand, the exact price or wording to show, and where the image should remain clear. Recommendations belong in context; reviewed spoken words can arrive as VTT. V2 recognises the source audio when no transcript is supplied; it preserves supplied lines instead of paying to recognise them again. The older directions do not transcribe audio. Turn off Burn in subtitles when the source already contains them.
Add a short title above the opening shot, then show subtitles only while the narrator speaks. Keep both clear of the gardener’s hands and the object being revealed.
Choose What ▶ does → Plan + decoration studio to inspect and edit the design before rendering. The default for a new V2 node is Assemble right on the server, so a connected chain can finish in one run.
Keep one visual series
Section titled “Keep one visual series”Open the library’s design systems tab to preview ten built-in systems available to every Studio account. Each contains reusable titles, subtitles, labels, callouts and calls to action, plus components suited to its purpose:
| Built-in system | What its graphics explain |
|---|---|
| TikTok · Pulse | An emphatic phrase, a numbered moment and energetic subtitle typography. |
| Editorial · Folio | A magazine composition with a framed video, an edition credit and an attributed quotation. |
| Signal · Detail | A visible object detail, a timed pointer and a named specification. |
| Kinetic · Poster | A large phrase revealed in two beats and a chapter ribbon. |
| Explainer · Flow | Three connected steps and a closer explanation of one step. |
| Data · Ledger | A number with its unit and source, or a comparison of two stated values. |
| Compare · Grid | Two approaches or sets of features, followed by a reasoned conclusion. |
| Timeline · Journey | An ordered sequence of events and an individual milestone. |
| Documentary · Notes | A person’s identity, a place and date, or an attributed quotation. |
| Product · Showcase | Product features, a pointer to a detail and an offer with its terms and action. |
The demonstrations use the complete, already-published 30-second AI short film The Key («Ключ»). Each shows that system’s own components and motion over the actual moving image, including its distinctive scheme, comparison or callout. Labels identify sample captions and editorial notes. Pointers mark a detail during a chosen interval; they do not establish automatic object tracking. V2’s reusable caption components preserve whole subtitle cues; per-word highlighting belongs to the previous mode described below.
The library cards immediately show a short description and a representative frame with that style’s graphics. Play starts the demonstration directly in the card; only the selected card plays. Choose На холст to add a Design system node, then connect it to Decoration. The node and library use the same pre-rendered demonstration. Previewing does not run a paid generation. Saved or imported editions without that exact demonstration use a live preview of their own unchanged components on the same full film.
V2 outputs the exact style used by its result through the same node’s дизайн-система (Design system) output. Connect that output to the next episode’s дизайн-система input, or save it to the library from the panel or Series design. This is an internal studio resource. New episodes bind new words, positions and intervals to the saved components and motion.
An agent working over MCP reads the same style back from the finished video: get_artifact returns design_profile for your own V2-decorated video. Pass it unchanged as series_profile to the next decorate_video, together with renderer: 'hyperframes' and decoration_version: 2; without them the style is not kept, and the response says so in series_profile_warning. The style is stored with the video, so it remains available when the original decorate_video response was lost.
To reuse the style in another project, choose the saved entry from the library’s design systems tab and add it to the canvas. Its components, motion and revision are preserved. Saving from Series design links the node to its library entry, so reopening the studio keeps that saved state. MD and TXT imports remain available as written design guidance.
A connected executable style lets the agent reuse its components instead of writing their design again. The new episode still needs analysis and placement decisions. This does not yet establish a particular speed or price reduction; the current per-minute decoration tariff remains the same.
You can also add Design system (Дизайн-система) with an existing library style or an MD document. It remains an optional source, rather than a required preparation step.
Choose ★ Save to library on Design system. In another project, select that entry from the library’s design systems tab and connect it to the new episode. The saved profile carries the same visual rules and pinned skill revisions. A recipe containing these nodes also carries that edition. Editing a library entry does not silently change nodes already using a saved copy; choose the updated system deliberately when you want another edition.
Choose стиль серии → создать новый стиль when an unconnected episode needs a different identity. A connected design system supplies the series identity. See skills and saved generators for the difference between a design system, an authoring skill and a saved node setup.
Review, edit and export
Section titled “Review, edit and export”In Series design, use the same playback, timeline zoom, snapping, undo/redo and fullscreen controls as the Assembly editor. The playhead and ruler use the composition’s own frame rate. Select an element in the picture or on its track to open its inspector.
- Correct every text field of the selected element. Moving focus out of a field applies the edit.
- Drag its timeline block to change when it appears; drag either edge to change its interval. Snapping helps align boundaries; hold Alt to bypass it temporarily.
- Drag the selection in the picture to move it, or use its width handle. The inspector also accepts its start, duration, X/Y position and width.
- Select a saved style component in Choose component, then use + to insert it, or remove the selected element. A subtitle component needs an existing source cue.
- Use ↶ and ↷ to undo or redo. A completed drag is one edit; the saved style stays unchanged.
New words, positions and intervals belong to this episode. The editor keeps the existing component templates, fonts and motion across the series, including when text was previously edited in the saved composition. Older custom HTML compositions retain their complete text and text-editing workflow; structural controls are available for compositions that can be rebuilt from their saved kit and element instances.
Choose Render · $0 to export the shown composition on the server through HyperFrames. Preview and export use the same saved composition. Editing text and rebuilding it do not call the authoring model again. Restore design returns to the generated composition; it asks before removing manual changes.
Creating a new composition is a paid operation. V2 has a duration-based price: the rate per minute is applied to the measured seconds, subject to the displayed minimum charge. An 80-second video is priced as 1⅓ minutes, rather than two started minutes. The card shows the rate, measured duration and total for your current pricing plan before the run; while duration is unknown it shows that it is being measured. The server verifies the source duration before reserving credits. A changed quote requires a fresh run rather than silently spending more. Analysis, authoring and the initial server render are included in that fixed successful-run price; the agent’s internal spending limit does not raise the bill. The older explicit directions retain their previous spending-ceiling and actual-usage pricing.
Source observations are privately cached by owner and file contents, independently of the design. A saved composition’s rebuild is free. Re-generating the design asks the model again, so review its price before replacing manual work.
Recover the same job and saved design
Section titled “Recover the same job and saved design”Decoration saves its job before analysis and authoring begin. Reloading the page or reopening the project restores the current phase, progress and any saved composition; it does not request a second design. Cloud rendering of the saved composition continues independently of the browser, and the Studio can recover its result after a server restart. Progress is not a guaranteed finish time.
For a full Assemble right on the server run, the quoted amount remains reserved until the finished result succeeds. A failed run releases its pending reservation and retains an already saved plan, so it can be inspected and technically rebuilt without authoring again. Plan + decoration studio is a successful paid operation once its plan has been reliably saved, even before a video is rendered; exporting that saved plan later remains $0.
If a server restart interrupts analysis or an unfinished authoring call before cloud rendering was submitted, the job reports the interruption and releases its pending reservation. It does not silently start another paid authoring call. Inspect the recovered job and saved plan before choosing to generate again. These lifecycle changes do not change the existing decoration tariffs.
Existing nodes keep their previous rendering mode until you explicitly change the direction. Its selector entry is Прежнее оформление. That mode retains its decoration studio, Build here browser export and On server export. The new directions export on the server.
Verify the result
Section titled “Verify the result”Read the entire exported video, including the last subtitle. Check spelling, timing, line breaks, contrast and cropped text on the intended screen shape. A background sign generated by the model is not a subtitle. Automatic recognition and generated lettering both need review; silence should not acquire invented captions.
TikTok look in the previous mode
Section titled “TikTok look in the previous mode”In Прежнее оформление, Subtitle look has two values, Card and TikTok. Russian label: Вид субтитров: Карточка / TikTok. Editorial, Signal and Kinetic use the subtitle treatment of their series instead of this separate selector.
- Card shows the whole subtitle line at once, as before.
- TikTok shows up to four words at a time in one or two lines, in UPPER CASE, Montserrat Black, white letters with a black outline. The word being spoken turns #FE2C55 (a fixed colour; there is no colour choice) and hops. Spoken and not-yet-spoken words stay white. The block never sits lower than 75% of the frame height, and a page breaks and fades out in a pause instead of hanging on screen. A long word is never cut: it is narrowed, wrapped at a hyphen or set smaller. The line text, start and end are the subtitle’s own; a page only shows a piece of the line.
New and old nodes. A newly created Decoration node switched to the previous mode uses TikTok. A node saved before the subtitle-look field existed keeps the Card look, including after 🔄; the planner never chooses the look, so re-planning does not change it. Choosing Card removes the field from the node, so such a node is the same as an old one. Opening a saved graph does not mark a finished Decoration stale because of the look field. A recipe carries the look as the author had it: a Card node, including one saved before the field existed, arrives as Card, and so does a Decoration node from a recipe published before this field. The copilot reads a node without the field as Card.
Switching the look costs $0. After you change only the look, ▶ does not ask the director again: it opens the studio with the same plan, and Build here (or On server) renders the video again at $0. This also holds when the plan was rebuilt after a finished video and the studio was closed: the look change is compared with that plan, not with the old video. The credit estimate shows $0, and the run summary and run history record $0 for such a server render. The preview in the studio and the mock-up in the node panel show the chosen look. In the TikTok look the card colour and background rows of the panel are replaced by a read-only note: white letters, black outline, #FE2C55 highlight.
Word timings. The Subtitles node sends the words to Decoration over the same VTT wire as WebVTT cue timestamps (<00:00:01.450>word). Downloaded SRT and VTT files, and any other text input that the VTT is connected to, get the plain text without those marks. A third-party VTT works too: tags such as <v>, <c>, <b>, <i>, <u>, <lang> and <ruby> and HTML entities are removed and are not burned into the picture. A VTT without word marks has no per-word timings. A project subtitled before the TikTok look sends the marks again when it is opened: the Subtitles node rebuilds its output from the saved recognition, so a TikTok Decoration highlights the words without a new recognition run. A Card node does not use the marks: its run signature and request stay as before.
Lines without word timings. A line edited by hand in the studio, a line whose text was normalised, and a line from a VTT without word marks are shown in the TikTok look without the highlight. This is visible: the studio warns before you edit a line that all lines become manual and the word highlight is switched off (restore it with Restore AI plan), and afterwards the plan summary says “word highlight is off for K of N captions” with the reason.
API. The decorate_video tool takes caption_look among its params: tiktok or card; omitted or unknown means Card.
When a subtitle appears
Section titled “When a subtitle appears”On a song, or speech over music, keep the Subtitles node on its default recognition model, Scribe v2 — songs and music: it builds the lines from individual words, and on the 18 October 2026 song where Whisper skipped the opening rap it recognised every line. Older nodes also use Scribe v2 on their next run; an existing result is not re-run by itself. See transcription for the fallback and what the status line says.
Each subtitle line starts when the first word of that line is spoken, not at the boundary of the recognised segment. Studio takes the later of the two (the segment start and the first word’s start) from the same recognition pass, so a line no longer shows up before the voice. The end time and the text are unchanged, and the exported SRT and VTT files start at the same moment. If a line cannot be matched to its words, it keeps its original timing; when that happens, the status line of the Subtitles node says how many lines were matched, for example word-timed 3 of 5. When every line is matched, the status line does not mention it.
Projects subtitled earlier keep their old timing until you re-run the Subtitles node with 🔄. Recognition is a paid operation; creating a new decoration plan is a separate paid operation. Review the current credit estimates before running them.
Availability and evidence
Section titled “Availability and evidence”This technique is documented as an authoring or workflow practice. It is available only through the stated inputs and operations; it is not a new model parameter. An unrendered example is untested. A dated successful case is an observation for its exact model, version, quality and mode, not a guarantee for another scene. Choose a model by the required operation first.
Related guides: transcription, composition, silence.
See the model–technique applicability matrix for connected operations, dated outcomes and configurations that remain untested.
Reviewed 21 October 2026 against the ten executable styles, full-film demonstration source, saved-profile and editor contracts. Focused checks cover all component slots and local font pins, library-to-node/V2 reuse, saved editions, complete manual text during structural edits, and the shared frame-rate-aware controls. Earlier browser checks covered Cyrillic text, local fonts, seeking, manual edits and undo/redo; the job-adoption check used a mock capture step. A separate short HyperFrames export was checked. These checks do not establish long-project readiness, a measured speed improvement or the quality of a newly authored design. The dated recognition observation above remains scoped to its original case.