Scene3D supports rigged characters, imported motion and editable timeline playback in the website editor and native Video Editor. This guide covers the built-in starting points, Meshy rigging, the free motion library and the boundaries of retargeting.
Start with a character
- Add object offers a Casual human (Meshy), an Expressive robot (Quaternius CC0) and a rigged mannequin.
- Tools → Characters & motion opens the character templates and saved characters/poses.
- You can import a project-owned, textured, self-contained rigged GLB. Raw Mixamo assets are not bundled; make sure imported models are yours to use.
When a scene contains one visible model, opening the full-screen editor selects it and displays its skeleton by default. Skeletons and helper controls are editing aids; they do not appear in final website/video renders.
Generate and rig a Meshy character
- Generate an eligible textured humanoid using Meshy V6 or V7 image-to-3D. Review the current generation quote shown for that choice.
- In Flowgen model settings choose Humanoid rigging · Meshy → Rig / resume humanoid. In a Scene3D model inspector, use Auto-rig with Meshy.
- Review the separate rigging quote before confirming. Rigging is currently a distinct 6-Flowie action, not included in model generation. Generation pricing varies: a tested V7 example quoted 71 Flowies, but the current UI quote—not that example—is authoritative.
- The panel supports stop and recovery/resume of the original accepted job. A completed rigged model is saved as a new version; the original remains in version history.
Text-preview outputs are not eligible for this rigging path until you create a textured model. Eligibility depends on the model and is shown by the current UI; do not assume every generated object is riggable.
Apply a motion from the free library
- Select a rigged character and open Properties → Motion capture → Free motion library.
- Search the 2,425 pinned CMU recordings, preview the skeleton, and set the starting time and clip length.
- Choose Use motion / review mapping. Inspect the suggested source-to-character bone mapping before you click Apply motion capture.
The selected recording loads directly; you do not need to download it manually. Meshy spine names that count downward are mapped by tracing actual joint ancestry, rather than assuming similarly numbered bones correspond. A tested Meshy sample matched 22 joints; that result is not a guarantee for every model.
Correct joints in the viewport
Click a skeleton joint to select its rotation gizmo and scroll the inspector to that joint’s Skeleton offsets row. Shortcuts cover Head, Left Hand, Right Hand, Left Foot and Right Foot. Add control creates per-axis XYZ rotation offsets that save with the character and layer additively on top of animation. Reset clears the correction. These controls change rotations—not rest-joint positions or provider skin weights—so inspect the character’s deformation after editing.
Use Edit in viewport to return to the editable scene if you maximized or swapped the camera preview. The timeline pauses while you adjust offsets.
Bind motion to the scene timeline
- Open Model Animation. An imported recording appears as the active Motion even when the GLB has no original animation clips.
- Choose Bind to timeline. Flowmo creates linear Playback keys from 0% to 100% across the motion duration, disables autoplay and opens the scene timeline.
- Rebind to timeline replaces the existing Playback track. Manual Playback scrubbing creates the channel automatically. The Playback readout follows timeline position; at 100% the character holds the last frame.
GLB clips and imported motion share the same animation target. A source clip with no reported duration uses a 5-second track that you can retime by moving its keys. Playback keys and the scene are saved for website/compositor use.
What retargeting does not do
- BVH transfer applies joint rotations in place. It does not transfer root travel, vertical jumps or foot-contact solving.
- Unmapped helper or finger joints may remain static.
- Joint offsets do not repaint provider skin weights; retargeting cannot guarantee skin deformation quality.
- The external MCP retarget tool accepts BVH text supplied by the caller. It does not search or download from the UI’s CMU motion library.
Using the native Scene3D agent
Inspect the loaded rig first. For caller-supplied BVH, call retarget_character_motion with action:"inspectMocap", review its proposed mapping, then apply action:"mocap" with the returned current scene hash, reviewed explicit mapping and inPlace:true. Do not invent MCP actions for the native Meshy UI or per-joint Skeleton offsets. For timeline work, inspect the live keyframe_scene declaration and exposed scene fields first; use only a Playback route that schema actually supports.
Video compositions
Connect the Flowgen Scene3D node’s scene_out or a 3D model output to the Video Editor node’s 3D Scene / Model input (scene_in). This creates an editable native 3D layer and retains the scene graph, materials, rig, motion and scene keys. Upstream updates preserve the authored layer’s placement and timing. Layer Input is for rendered video, not an editable 3D source.
Verified behavior
Recent UI QA exercised a Meshy character with imported motion, the default selected model/skeleton on reopen, joint click-through and foot offsets. Playback was sampled from the timeline; pausing and backward seeking were deterministic, and resetting the test foot offset restored its original position. The character and extruded text were also viewed together in website and native video composition previews. This verifies those tested routes—not every model or rig.