Generate 3D Scene

The Generate 3D Scene node turns a description of a shot into an editable, animated 3D scene in untextured clay. It places the objects, the camera and their motion over time, so you can check framing, camera moves and blocking before you spend credits on a final video. The node returns the scene, not a video: edit it with Edit 3D Scene, or export it with Render Video.

When to use it

• You want to previsualize a shot: where the subjects stand, what is in front of what, and how the camera moves.
• You want a layout reference that makes a video model follow your blocking and camera move.
• You want to rebuild the rough layout or motion of a reference image or clip, then adjust it.
• You want to compare camera moves and timings cheaply before a final render.

Where it is available

Generate 3D Scene is in the node picker on every edition, under Video › Titles, Graphics & Captions. Code and AI assistants can run it through the API, the SDK and the generate_3d_scene MCP tool. On self-hosted installs there is no credit billing.

Quick start

Inputs

The output, Composition, is the editable scene. Connect it to Render Video, or to the Scene input of Edit 3D Scene or 3D Render Pro.

Settings

A role tells the model what a reference means: what things look like, where things go, or how they move. The same picture means three different things under the three roles.

The preview and revisions

The scene stores every object's id, position, rotation, scale and size, the camera's position, target and lens, the lighting and the keyframes. The 3D preview in the settings panel lets you play and scrub the scene, select objects and edit their values directly.

• Direct edits are free. Changing a value in the preview creates a new revision of the scene, without any model call and without credits.
• Animated or not. On a channel that is not animated, an edit moves the object at every frame. On an animated channel, an edit writes a keyframe at the current frame.
• Every revision is kept. The revision list shows where each one came from: Generated, Model edit, Manual edit or From the connected scene. Restore any revision to make it active again.
• Your edits win. If a generation finishes after you edited the scene, your edits stay active. A notice offers Use the new one or Keep mine.
• Units. Positions are in meters with Y pointing up, rotations are in radians, and frames count from 0.

Each MP4 you render uses one specific revision.

What the scene can and cannot do

• Simple geometry. The Basic engine builds scenes from simple shapes and groups, including simple stand-in figures for people. It does not build detailed textured models or physics simulations.
• Approximate reconstruction. An image cannot show what is hidden in it, so hidden geometry is guessed. A video reference is read as a guide for motion and layout.
• Whole clips only. A video reference is analyzed as a whole. To use part of a clip, trim it first with Trim Video and reference the trimmed clip.

Always check the preview before you export.

Use the render as a video reference

The MP4 that Render Video exports from a scene is a layout reference. It carries where the subjects are, what is in front of what, the framing, the camera move and the timing. It also carries a look, untextured grey clay, and a video model copies that look unless it is told not to. Two rules keep the layout and drop the clay.

1. Never attach the clay render without a scoping line. Wire the Render Video output into the Video Refs input of a Generate Video node, and Nodaro adds the line for that reference by itself. It does so on a workflow run and on the video node's own Run. The node's Final prompt preview shows the line exactly as it is sent. The line for a clip is:

LAYOUT reference only — match its subject positions and blocking, its foreground occlusion, its framing, its camera angle, its camera motion and its timing. Ignore its untextured grey clay placeholder look, its flat placeholder colours, its materials, its lighting and its empty background; none of that is the target look. Take the look from the prompt and from the other references

The model reads it as @video_1: followed by that line. For a scene whose camera does not move, the line leaves out the camera-motion clause. For a frame extracted from the render and wired as an image reference, it leaves out the motion and timing clauses. A prompt that already scopes that reference is left alone, so running again never doubles the line.

2. Give every figure that must look real its own character reference. Photoreal treatment follows each referenced subject, not the whole frame. With one layout reference and one character, only that character becomes real, and every other figure stays a clay stand-in. With one Character Asset per figure, every figure becomes real and keeps its identity. Keep two reference slots free for a location or a style image.

Also connect your original appearance images to the final video node: the clay render supplies the blocking and the camera, and the images supply the look. Wire the render into a reference input, never into the start frame, because a start frame sets the look and no scoping line reaches it. To compare a guided and an unguided result, keep the prompt, the images, the model and the settings identical, and add only the clay video with its line.

Credits

On Nodaro Cloud, a scene costs the sum of up to three parts:

Authoring is one call to the AI model. The call is charged for the tokens it actually used, at the selected model's rate. The tokens are your prompt, the descriptions of the references, any video analysis text, and the scene plan the model writes. The full rules are in How LLM calls are charged.

• Before the run, the Run button shows an estimate (≈ and a number) of a typical scene on the selected model. Gemini 3.6 Flash costs less than the default, Claude Sonnet 5.5, and Claude Opus 5.5 costs more.
• A longer job costs more. A long prompt, many references or a repair pass use more tokens.
• Reasoning Effort has no price of its own. On a Claude model, any effort runs the call on Anthropic's own service, at that service's rate.
• Failures. A failure that the model's maker billed is charged for that usage. A failure that was not billed is refunded.

Every aspect ratio this node offers renders at 1920 pixels or less, so its exports cost 55 credits. A larger frame appears only when a scene is resized in the editor, or supplied through the API or MCP. See What a 3D scene render costs.

Playing the preview and editing values are free. Community and Business editions do not use credit billing.

Tips

• Say what the camera does and when. "Dolly right over four seconds" gives the model a clear move and duration.
• Keep previz short. The default of 4 seconds at 24 fps is one beat of blocking, which is what a previz pass is for.
• Lock what must not change. Before you ask Edit 3D Scene for a change, lock the objects that must stay as they are.
• Need a finished shot in one step? 3D Render Pro authors and exports a more detailed scene in one run, where your install offers it.

Troubleshooting

The preview says it needs WebGL. The 3D preview needs WebGL, and the browser has it turned off. Enable WebGL in the browser, or use a browser that supports it. The scene itself is intact: you can still render it to video.

The preview stops or loses its graphics context. The scene data is not affected. Keep editing values in the panel, or render the scene to video.

From the API

POST /v1/3d-scene/generate returns a jobId. When the job completes, its output_data.scenePlan holds the editable scene. Render it with POST /v1/render-video/plan and planType: "3d-scene".

const scene = await client.nodes.runAndWait("generate-3d-scene", {
  prompt: "A red suitcase rolls behind a central pillar and reappears. Dolly right over four seconds.",
  durationSeconds: 4,
  fps: 24,
  aspectRatio: "16:9",
  references: [{ id: "suitcase-appearance", kind: "image", role: "appearance", url: appearanceImageUrl }],
});
const video = await client.nodes.runAndWait("render-video", { planType: "3d-scene", plan: scene.scenePlan });

• References use { id, url, kind, role }: kind is image or video, and role is appearance, layout or motion. An optional objectId ties a reference to one object. Partial time windows of a video are refused before any charge.
• Frame rate. Through the API, fps can be any value from 15 to 60.
• Engines. GET /v1/3d-scene/capabilities lists the optional advanced engines of the install. Asking for an engine that is not available is refused; it never falls back to Basic. With an advanced engine that supports imports, inputAssets can bring in up to 8 existing GLB models that your scenes retained.
• Advanced results. A scene from an advanced engine also reports the assumptions it made and the repairs it ran. It can be delivered without the visual review's approval, and a failed run can keep the draft it built. 3D Render Pro explains how to read these results.

AI assistants use the generate_3d_scene, edit_3d_scene and render_3d_scene MCP tools. See 3D scenes through MCP and the 3D scenes API.