Elements, scenes and media

An Element describes something you want to carry across scenes. A scene describes what happens in one part of the film. Media is the image, video or audio you generate or bring into the project.

Elements carry identity and context

An Element (Элемент) can represent a character, object, location or style sample. Its name, role, description and reference material help the writing and generation steps use it consistently.

For example, a recurring detective belongs in an Element. The detective finding a letter belongs in a scene. The resulting eight-second clip is media.

A reusable library Element is applied to an Element node, then connected to the nodes that need it. It is not applied directly to every node with an Element input. Through MCP, the catalog identifies nodes that accept library Elements.

Reference material helps guide a model; it does not guarantee an identical appearance in every result. Check the generated frames and clips.

Scenes carry the action

Prompt Scriptwriter (Промпт Сценарист) turns the idea and connected Elements into scenes with image and video prompts. A scene can travel through several nodes, so the graph can show both the scene arriving at a node and the scene leaving it.

When inspecting a graph through MCP, scene_in and scene_out identify these scene containers. describe_scene reads one of those returned scene addresses.

Expanding written scenes builds their image and video branches. The studio chooses an applicable approach, such as a first frame or a storyboard sheet. Expansion prepares nodes; the clips still need to be generated.

The parameter pill below Video shows the model, actual aspect ratio, resolution, quality, duration and sound in the same order as the image controls. Click an available setting to open its editor; the panel and pill edit the same node settings. The aspect ratio is a read-only report of the resolved result format. A speaker or mute icon shows the current sound state; models without a sound-off option keep that control read-only. On a narrow canvas or at a larger zoom, the pill wraps so its settings and actions remain reachable.

Scenes on song timecodes

For a music video, paste the song’s subtitles (SRT or WebVTT) into the brief of Prompt Scriptwriter or Stager (Постановщик), say that you want a clip to this song and give its length in your own words, for example song length 80 s, песня длина 80с or трек идёт минуту двадцать. The studio finds the subtitle lines itself; the writer reads the brief, decides whether it asks for a video to this song and returns the song length as a number of seconds. For Stager only the brief section of its task counts, so times in a concept or promise written earlier in the graph do not count as subtitles.

When the writer ties the scenes to the song, it only chooses where each scene starts. The studio checks those starts and calculates each scene’s on-screen plan: the next scene’s start minus its own, with the last scene ending with the song, so the plans add up to the song length exactly. When the brief gives no song length, the video ends with the last subtitle line and a visible line says the outro has no scene. A scene never starts inside a sung line; it starts at a line start, in a pause or where two lines meet. The Scriptwriter or Stager card then shows a chip with a clock icon at the bottom such as by timecodes · 10 scenes = 80.0 s, the top of its panel shows the full line by timecodes: 10 scenes = 80.0 s (length from the brief), and each scene carries its plan to Video and Montage. The Director window shows the same plan: the header reads 10 scenes · 1:20 by timecodes · shooting 1:23, each scene row shows its plan length with shooting N beside it when the clip is longer, and under the scene’s seconds stands the same line as on Video. A clip is generated for the plan rounded up to whole seconds within the model’s limits; Montage trims the extra tail, see assemble and finish.

The studio refuses instead of guessing, and names the scene or line with its numbers:

• Nothing is refused before the writer is called: subtitles in the text are not yet a clip to the song (a transcript pasted as material is fine), and the writer decides that by tying the scenes.
• After the writer ties the scenes, from the subtitles: one sung stretch is longer than the longest clip of the selected video model (30 s when no model is known), or a timing line is broken. Fix the subtitles and run again.
• After the writer ties the scenes, from the answer: the song length is shorter than the end of the last line, some scenes have a start and others do not, a first scene not at 0, starts that do not rise, a scene starting inside a line, a last scene starting at the end of the song or later (when no length is given, at the end of the last line or later — then name the song length in the brief), a scene shorter than 1 s or longer than the model’s longest clip. Run again for a new answer, or fix the timecodes or song length in the brief.
• Either way the charge is refunded and nothing is cached.

Without SRT or WebVTT lines the request to the writer is unchanged; subtitles in another format (SubViewer, ASS, MicroDVD, LRC) are not read as timecodes, so save them as .srt. When the writer does not tie the scenes to the song, the scenes keep their usual lengths, the card shows the chip timecodes off — scenes not tied, and a visible line says so and asks you to name the clip and the song length in the brief. Timing lines that could not be parsed are counted and named.

On Video a scene with a plan shows by timecodes 8.0 s under the length in the panel’s Next run block, and the length chip on the pill shows a clock icon with 8.0 s (or 5.28 → 6 s when the clip is generated longer; ⚠ 10.28 → 10 s when shorter), with the full line as its tooltip. This holds before the storyboard sheet or frame above it is drawn: the plan is read from the scene that feeds that picture. When the model shoots longer, the line names the tail and whether Montage will cut it. When the clip would be shorter than the plan — for example after switching to a model with a shorter limit, or after setting your own length — a warning line says how much earlier the clip ends than its timecode and what to change, before you pay. The writer’s choice of boundaries can vary between runs; the checks above apply to every answer.

Media is the material you can use

An image may be a character reference, a scene’s first frame or a sheet of planned shots. Those uses give it different jobs even though the file is still an image.

Generated media and uploaded files appear in the material library. MCP can list materials and request a download link for one file. Documents can also be stored there, but storing a document does not automatically extract its text or turn it into an image reference.

A song can also come from a Spotify track link: paste it on the canvas, or press Spotify on an Audio · file node. Spotify does not hand out full tracks, so the studio reads only the title, artists and length from the track page and imports the same recording’s audio from YouTube Music. It accepts a recording only when the artist and the length (within 2 seconds) match. A slowed, live or remixed version has a different length and is refused rather than substituted. After the import, a notice names the YouTube Music address the audio came from. When the full recording is not available to the studio server, the node gets Spotify’s own preview instead: a short MP3 fragment whose length Spotify chooses (often well under 30 seconds). The node name ends with preview N s, and the notice says how much of the song it is; upload the song file to analyse the whole song. Album and playlist links are refused with the reason. A track without a preview whose recording cannot be found is refused too; when YouTube refuses the studio server’s requests, the error says that YouTube blocked the request rather than that the track was not found.

Read frames, sheets and keyframes before choosing an image’s role, and instructions and results to follow current scene text. For a practical example, follow a consistent character.

Last checked: 18 October 2026. The exact media inputs accepted by a video model are part of its current catalog entry.