3D scenes & animation

The workspace combines a viewport, object inspector and animation timeline on one page.

Scenes and views#

Open 3D Scenes, create a scene or open the three-hall sample. The library supports context menus, multiple selection and batch deletion. Each card's thumbnail is a top-down plan drawn from the scene data itself: every object paints its own colour where it meets the floor, cylinders and spheres draw as ellipses, and a camera move draws as a dashed path. It follows edits — it is not a screenshot frozen at the last save. Switching to another workspace page and returning preserves the open scene. Click its title to rename it in a dialog.

• Free view: orbit, pan and zoom while positioning objects. Moving this observation camera does not insert animation keys.
• Camera view: see the current shot's actual composition and aspect ratio.
• Overview: observe the moving camera and its path from an independent view, alongside a live shot inset.

Views are remembered per scene: reopening a scene returns you to the view mode, shot and camera position you left it at. This lives only on your device and never changes the scene or syncs to collaborators.

The ground grid has no edge: cell size steps with camera distance (1 → 10 → 100 m) and the fade radius scales with it, so the grid is still there hundreds of metres out; the X axis is tinted red and the Z axis blue for orientation. An axis gizmo in the viewport's top-right corner turns with the view in real time; click any of its balls to swing the camera onto that axis (from above, from the front, and so on) without changing distance. The toolbar provides move, rotate, scale, snapping, view presets and fullscreen. Trackpad/mouse preferences are shared with other canvases. F focuses the view. Fullscreen keeps only the 3D view and the viewport's own toolbar — the page header, timeline, inspector and agent pane are hidden; Escape exits. The viewport is not remounted, so the model, camera and any running animation stay exactly where they were.

Objects and lighting#

Objects in scene is a tree: groups can nest without a depth limit, children sit under their parent and are indented by level, and any row with children collapses from its chevron (collapsed rows show how many items they hold). Icons distinguish groups, cameras, lights and geometry. To put an object in a group, pick it under Group — the dropdown shows the full path (for example Outer / Inner), because new groups are all named Group.

The ＋ beside Objects in scene opens a searchable, grouped menu of primitives, rooms, stairs, tables, figures, lights, groups and model import. Import has one entry in this menu.

Use the gizmo or inspector to edit position, rotation and scale. Units are meters and degrees; snapping uses 0.25 meters, 15 degrees and 0.1 scale steps. Edit object materials and scene background, ambient light, key-light direction, intensity, temperature and softness.

Whole sidebar sections can collapse; properties inside a section remain visible. Drag the panel boundaries to resize them, including the left edge of the inspector and of the docked modeling assistant; widths are remembered. Objects can be copied, hidden, grouped and deleted. Deleting a group deletes its descendants. Deleting a camera that a shot uses (or a group containing it) removes the shots using it as well, with a notice naming them; only the camera of the scene's last shot is refused, because a scene needs at least one shot.

Hiding and showing#

Each row in the object list has an eye button next to delete that hides or shows the object; H toggles the selection and ⌥H shows everything (as in Blender). Hiding is undoable and autosaves.

• Hiding a group hides everything in it; showing the group again restores whatever each child's own state was. Rows you hid yourself are dimmed and italic; rows hidden only because their group is are just dimmed.
• Hidden objects are invisible in the viewport and cannot be clicked there. They are left out of exported frames, camera videos and GLB files, out of the object list handed to generation models, and off the scene card's thumbnail.
• A hidden camera still works as a shot's camera — the shot reads the camera's data, not the camera model drawn in the viewport.

The modeling assistant#

The assistant on the right builds the scene with the chat model you selected: describe the space and the objects, and it edits this scene directly. After each step it renders the scene as a few blockout images and looks at them — the shot camera's framing plus an overview, a top view and front/side elevations — to catch objects sunk into the floor, floating, overlapping or blocking a doorway, none of which show up in the numbers. Those images are used only within that turn and never enter your media library; models without vision skip this step.

For shapes the primitives cannot express — architectural detail, props, machinery — it can model directly in your Blender; see Models and Blender below.

One timeline for every object#

Every object has a row, including objects with no animation. A camera is an object too; the shot selector chooses which camera to use, its duration and aspect ratio.

1. Select an object in the viewport, object list or timeline.
2. Move the playhead to a time and adjust its pose.
3. Click Insert keyframe or press I to record position and rotation/scale, or camera target/FOV.
4. Move to another time, adjust the pose, insert another key and play the result.

An empty track stays still. Changing a pose does not automatically create a key. Objects and cameras share this workflow; there is no separate movement recorder.

The single toolbar contains shot selection, add/delete shot, playback, time, insert/delete keys, filtering and zoom. ＋ creates a shot along with a new camera at your current view; − next to it deletes the current shot together with its camera, unless another shot still uses that camera. The last shot cannot be deleted. The filter menu provides object search and animated-only filtering. Shift selects multiple keys; drag moves them and Delete removes them. Cmd/Ctrl+Z undoes, Cmd/Ctrl+Shift+Z redoes. Tracks support up to 100 keys; shots up to 120 seconds.

Generate images or videos#

Generate on the toolbar no longer exports a frame first: it opens a creative board with this scene wired into a generation item, where you choose an image or video model, describe the look and generate. The three ways in are "Generate an image from this shot's composition", "Generate from first and last frames" and "Generate from the camera-move video". The shot is rendered from the scene as it is at that moment — change the scene and generate again, and the new version is what gets used; an unsaved scene asks you to save first (the server renders the saved version). The current lighting is written into the prompt; camera helpers never appear in reference frames, videos or GLB exports.

A figure in the scene can be told which character it Plays — any person in the asset library, picked in the inspector. When a shot shows those figures, generation brings along their looks and descriptions too (see Asset library). Scenes built from a script on a board get this filled in automatically.

On a creative board, a 3D scene item renders its own blockout references: select the scene item, pick a shot and what to render, and the first/last frame and camera-move video land to its right (see Creative boards). In workflows it is the "Render blockout references" node, with the scene and shot picked from dropdowns (a scene with a single shot needs no pick). Both use the same renderer: the backend has its own blockout renderer for the reference frames and the camera-move video, and what it draws — geometry, camera interpolation, light direction — matches what this page shows. Imported models are drawn too, using their own material colours. Anything it cannot draw — a Draco- or meshopt-compressed mesh, a model over the reference frame's triangle budget, a missing file — is reported one by one with the reason, rather than quietly leaving a gap.

A board's 3D scene item can also Build a 3D scene from text: connect a script or storyboard (document or note) and it builds a blockout set — figures move as written, several cameras cover the same action from different positions with camera moves, and each figure's Plays is filled in. Workflows have a node of the same name. The built scene lives on this page, ready for you to adjust the cameras.

Export saves frames, camera previews and models. Preview encoding uses deterministic MP4 frames where supported, with a WebM recording fallback when necessary.

Models and Blender#

Imported models belong to the workspace, not to one scene: import once and place the same prop in any scene in that workspace — the + menu lists them under Model library. Deleting a scene does not delete its models; deleting a model itself is refused while a scene still places it, and the refusal names those scenes.

Import self-contained GLB 2.0 files up to 512 MB or embedded-resource glTF 2.0 up to 100 MB. Draco, KTX2 and meshopt are supported. Real-time limits are 5,000 nodes and 2,000 meshes. External URLs and filesystem references inside models are rejected.

Install and enable the local Blender MCP connection using the plugin guide. Blender in the toolbar checks the connection, sends a saved scene and its camera tracks (imported models placed in the scene go along), receives changes and downloads the associated .blend file. Receiving replaces the current scene as one undoable edit, or creates a separate scene if requested. Get from Blender in the library turns the currently open Blender scene into a new scene without a previous send: the geometry becomes one model, and native cameras and lights come along.

• Cameras: each Blender camera becomes a camera object and a shot with the same name. Field of view follows Blender's sensor-fit rules, and the aspect ratio snaps to the nearest of 16:9, 9:16 or 1:1; animated cameras are sampled frame by frame (up to 100 frames and 120 seconds). A Mosael camera also needs a "looking at" distance, which only sets the orbit pivot and the interpolation between keys — it does not change the picture — so it is inferred from the camera's track target, depth-of-field focus and similar cues. Orthographic or panoramic cameras, and cameras that are rolled or look straight up or down, are not brought back, each with its reason. If no camera comes back, the new scene keeps its default camera.
• Lights: point lights come back as they are; spot and area lights come back as point lights, without their cone, size or direction; a sun becomes the scene's key light (azimuth, elevation and the closest colour temperature), and with several suns the brightest is used. Intensity conversion is the exact inverse of the send path, so a light sent and brought back keeps its strength.

Received geometry is one model, not an object-by-object merge. Associated cameras are sampled to at most 100 keys with linear timing. Keep advanced materials, skeletal animation, physics and native modifiers in the Blender project. The notices shown when getting or receiving from Blender follow the interface language. Mosael does not provide sculpting, Boolean modeling or a background Cycles rendering queue.

The modeling assistant can also work in Blender directly: it sends the current scene across (the same path the toolbar button uses, so no browser page has to be open), reads the scene, renders views to look at, runs modeling code, and brings the result back as one model in the current scene. Looking is not limited to the six presets: it can ask for any azimuth and elevation, move the camera in, frame only the part it just built, and switch to see-through shading to check whether two parts actually intersect — at whole-scene framing a seam is a few pixels wide, and that looks exactly like a part that was never built. Running code asks for your approval and pushes an undo step first, so you can ⌘Z it in Blender; if approving every step of a long modeling session gets tedious, the autopilot settings have a separate gate for Blender modeling.

Saving#

Edits autosave with revision history. A save conflict does not silently overwrite another draft. Restoring a revision creates a new revision. Scene, model and board references stay scoped to their workspace.