Storyboard-Level Control Over AI Video: Segments, Frames and Edit Mode

Any Nureta video longer than 15 seconds is rendered as a stitched multi-shot scene. The question is who writes the shot list. The API has three answers. Auto mode, the default: a built-in director plans and renders everything from one prompt. Advanced mode: you pass your own segments[] storyboard and it renders immediately. Edit mode: the director (or your segments) drafts a storyboard, the task parks at status planned, and nothing renders until you approve it. All three apply to scenes of 16–180 seconds.

Control matters most in adult storytelling because pacing is the product. A scene needs build-up before payoff, the same characters in every shot, and wardrobe and setting that don't drift. Auto mode handles that well for most prompts; the other two modes exist for studios and apps that need predictable, repeatable results, such as a series with a fixed character or a format where the story beats must land at specific seconds.

In advanced mode you send a segments array alongside the usual text item. Each segment is one stitched clip with a script, an optional duration_seconds from 4 to 15 (default 9), an optional first_frame and last_frame, and an explicit flag. The segments must total 16–180 seconds, and that total becomes the video's duration and price. For example: "segments":[{"script":"she steps onto the rooftop, city lights behind her","duration_seconds":8,"explicit":false},{"script":"he joins her at the railing; they kiss","duration_seconds":10,"explicit":false},{"script":"the payoff shot","duration_seconds":12,"explicit":true}].

Write each segment as a single visual beat: who is on screen, what changes, and where the camera is. Short, concrete scripts render more reliably than paragraphs, because each clip only has a few seconds to show one idea. If you already generate scripts with the API's $0.01 seahorse-script task, each returned scene (setting, beat, durationSeconds) maps almost one to one onto a segment.

The explicit flag answers one question per shot: does a sex act actually happen in this shot? The video model has no built-in knowledge of anatomy, so shots marked explicit: true are routed through a separate anatomy-reference pipeline. For that to work, the shot's script must describe the action concretely; a vague line such as "they get closer" on an explicit shot attaches no references and renders soft.

Shots marked explicit: false are build-up or story shots, such as a look, a touch, clothes coming off or afterglow, and are rendered from text only, with no anatomy references attached. This is what keeps a scene paced like a film: the setup stays suggestive and the payoff is explicit, instead of every shot averaging into the same register. In edit mode the director sets the flag for you, and you can override it shot by shot.

Pinning a first_frame or last_frame renders that shot frame-conditioned: the image becomes the opening or closing picture, which is useful for matching a thumbnail, continuing from a previous clip, or landing on an exact final composition. Frames can be set on a single clip through the role field on an image, on individual segments, or in the edit-mode plan.

A pinned frame replaces that shot's character and background reference; the two cannot be combined. The shot's identity then comes from the frame itself, so either pin frames that already show your character, or leave frames unset and let the shared character reference keep every shot consistent. The same rule applies to single clips: send a frame or reference images (up to 9), not both.

Submit with "mode":"edit" and a duration over 15 seconds. The task is charged at submit, drafts a storyboard and parks at planned. The scene field then holds a segmentCount, a characterImage preview, and segments with an index, script, durationSeconds, explicit flag, a people-free backgroundImage preview and any frames. If you set a callback_url, a best-effort draft-ready event fires once the draft is planned.

Two endpoints edit the draft. PUT …/tasks/:id/plan changes raw fields per index (script, duration, frames; null clears a frame). PUT …/tasks/:id/plan/structured is the recommended path: you edit each shot's short beat and the server re-expands it, the segments array is authoritative so you can reorder, add or remove shots, each shot can take a background_image plate, and top-level world, wardrobe and character override the shared style. Re-expansion is asynchronous: the call returns 202, and you poll until the scene is planned again with reexpandStatus ready.

Edits never reprice. The paid total is locked as scene.lockedDurationSeconds, and both edits and render reject a plan whose shot durations drift from it, so trimming one shot means lengthening another. When the plan is right, POST …/tasks/:id/render starts the render. If you change your mind first, POST …/tasks/:id/cancel refunds the full charge immediately and ends the task as cancelled. Drafts you never render are refunded automatically, and failed renders are refunded too.

Pricing is per second from a prepaid balance: $0.1888 per second at 480p, $0.2773 at 720p and $0.5428 at 1080p, so a 60-second storyboard at 480p costs about $11.33 however many times you edit it. The Nureta app exposes the same storyboard editor to individual users (per-shot script, duration, frames, background, shared World, Wardrobe and Character, and a per-shot toggle for explicit shots), but app pricing is different: $1.29 per 5-second 480p video, paid in USDC.