Dance in its World: Methods

1 · Problem statement

What the pipeline has to produce.

The task is the following. Given a motion-capture recording of a culturally specific dance, produce a costumed, world-situated embodiment of that recording that satisfies four constraints at once. The first is fidelity: the movement signal is preserved exactly, and nothing downstream alters the recorded motion. The second is de-identification: the embodiment carries no individual identity. The third is portability: the result renders in a browser. The fourth is provenance: every element that is generated rather than recorded is labeled as generated and is never presented as documentary fact.

De-identification is treated here as a design constraint with a precise definition. A de-identified embodiment contains no scan of a real body, no facial geometry tied to a person, and no biometric measurement. The body that carries the motion is parametric, which means it is generated from a set of numerical parameters rather than measured from anyone. The rest of this account is organized around how each stage either honors that constraint or marks where the work to honor it fully is still ahead.

2 · Input data — the motion capture

The recorded signal.

The recording is cataloged in the platform manifest as an Efik Ekombi capture performed by Sinclair, under the Atomic Confluences identifier ACF-0003.01, with its source format recorded as FBX.1 REAL The form the build actually processes is a derived JSON time series, and that is the file the rest of the pipeline reads.2 REAL Its specification is below.

The capture convention upstream of this file is the MediaPipe thirty-three-landmark full-body pose, which the platform normalizes and reduces to its stored joint skeleton, and in that reduction the face landmarks around the eyes, ears, and mouth are intentionally dropped.3 REAL Dropping the face at capture is the first de-identification step, and it happens before any body is ever drawn.

The stored skeleton is a kinematic tree. Each joint is a node with a parent, and each joint holds a local transform relative to that parent, while the root joint carries the global translation of the whole body through space. Posing the figure for a given frame is forward kinematics, which computes a joint’s placement in the world by composing the transforms along the path from the root down to that joint.

One note on the signal. The JSON stores joint positions, and the animation rig reconstructs joint rotations from those positions by aiming each bone down the chain. The recorded quantity is therefore position, and orientation is derived. The repository also carries a general conversion path for the cataloged formats, a Blender script that imports an FBX file and exports a BVH hierarchy, where the BVH form keeps the skeleton hierarchy and the per-frame rotations and discards the mesh and skin of the original.4 REAL For the Ekombi recording specifically, only the JSON form is present in the repository, so the methods here describe the JSON path that produced the figures.

The orientation convention is worth stating because it recurs in every export. The data is Y-up. The Blender step maps each point from (x, y, z) to (x, −z, y) so that Blender’s Z-up world, exported back with the glTF Y-up option, returns the original orientation the studio expects. REAL

3 · The body-representation ladder

One motion, five readings.

The same recording is carried by five increasingly embodied representations. Each is a real asset in the repository and an output of its own Blender script, and each preserves the motion while changing what surface the motion wears. The screenshots below are renders of those exact assets.

The skeleton is the joint set placed by forward kinematics, drawn as spheres at the joints and capsules along the bones, with no surface. The mannequin parents primitive volumes to the bones, so it inherits each bone’s transform and gains volume, but the volumes meet in visible seams and do not deform continuously. The skinned surface replaces those parts with a single continuous mesh whose vertices are weighted to the skeleton, and that is where real skinning mathematics enters.

Skinning is the standard method by which a single mesh follows a skeleton, often called linear blend skinning. Every vertex is assigned a small set of influencing bones with weights that sum to one, and the deformed position of a vertex is the weighted sum of where each influencing bone would carry it from rest. This is the deformation the glTF skin encodes and the browser plays back.5 REAL

The anatomical representation is generated by MakeHuman through its Blender successor MPFB, which produces a parametric human whose output is released under the CC0 public-domain dedication.6 GENERATEDParametric means the body is a fixed base topology deformed by a set of scalar parameters, the morph targets, so the figure is a point in a parameter space rather than a measurement of anyone. This is the load-bearing claim of the whole pipeline, and it is developed formally in the next section. The export is reduced to the body mesh alone, with the toolkit’s fitting-cage proxies removed by selecting the body vertex group and deleting the rest, and the head is then neutralized to a featureless form, so the figure is recognizably a human body and no individual’s likeness.REAL Driving it with the recorded motion, by retargeting, is the step still ahead for this rung. FORTHCOMING

The costumed figure is the skinned body dressed in the modeled Ekombi regalia, which is the subject of the arteme section below. Across all five rungs the motion is the one constant, and what changes is only the surface that wears it.

4 · De-identification, stated formally

Identity removed by construction.

The pipeline never contains a biometric of a real person. The movement is a sequence of joint coordinates, which is kinematic information about how a body moved and not a measurement that identifies whose body it was. The body that carries the movement is parametric, which makes it a vector in a morph-parameter space rather than a scan. The face, which is the part of the body most capable of recognition, is dropped at capture and absent from the procedural body. Each of those three properties is a guarantee about what the data is, rather than a filter applied after the fact.

This is the same ethic that governs the facial primitive in the wider grammar, the opseme, which is built for the preservation of trained facial movement and explicitly not for the identification of persons. Here that ethic is stated as a property of the data. The procedural skinned and costumed bodies carry a featureless head by construction, and the anatomical MakeHuman rung is now reduced to its body mesh with its head neutralized to a featureless form as well, so the by-construction guarantee holds across the bodies alike. REAL What remains ahead for the anatomical rung is driving it with the recorded motion. FORTHCOMING

5 · Costume and extension modeling — the arteme layer

Regalia bound to the kinematic chain.

The arteme is the material-cultural layer the body wears, and in computational terms it is deformable geometry bound to the same skeleton that carries the motion. The Ekombi regalia was modeled in Blender as separate mesh objects, one per element, and every element was weighted to the bone-aim armature so that it inherits the motion of the limb it sits on. REAL The modeled elements are a fitted blouse with short sleeve caps, a draped wrapper gathered at a beaded waistband, the banded raffia cuffs at the arms and the calves, a neck collar, beads at the waist and the wrists and the ankles, and a stacked crown.

The limit here is the deformation model. The regalia is rigid weighted geometry, which means it follows the bones through skinning but does not yet simulate cloth, so the wrapper swings with the leg rather than settling and folding as fabric would under gravity and contact. Cloth simulation for the regalia is the intended next step. FORTHCOMING

6 · World generation — the tereme and scapeme layer

A generated world, and a real-data seam.

The environment the dance stands in is produced by World Labs through its Marble world model, which generates a persistent three-dimensional scene from a text prompt. For the Ekombi build a scene was generated on the twenty-fourth of June 2026 under the model marble-1.1, and its operation and world identifiers, along with the returned assets, are archived in the repository.7 GENERATED The returned assets are gaussian-splat files at several densities, a collider mesh as a glTF binary, and an equirectangular panorama, and the scene carries a metric scale factor and a ground-plane offset for compositing a figure into it at the right size.

In the current browser cut the panorama is used as an equirectangular backdrop behind the figure, which is why the viewer’s In World mode maps that image onto the scene environment. Walking inside the world as a navigable space, using those gaussian splats, now runs live below: the scene is rendered in the browser on our own pipeline through Spark, World Labs’ open-source splat renderer, so the world stays on our site with no jump out. GENERATEDA second generation path, Higgsfield, is wired into the platform for fast look generation from a prompt that names a real place’s character rather than its coordinates, and it is available to the pipeline, though no Higgsfield image was produced for this particular build. FORTHCOMING

Kept deliberately separate from the generated world is a real-data seam. NASA’s Global Imagery Browse Services serve open, public-domain satellite imagery through a tiled web map service, and a MODIS true-color tile over the Cross River region of southeastern Nigeria, the region Ekombi comes from, was retrieved and saved as a layer in the build.8 REAL The service needs no key. The reason the two are kept apart is a matter of epistemics. The generated world is an experience, and it is never presented as the documented shrine, while the satellite layer is observed ground truth about the actual coastline and vegetation of the region. The first is labeled generated and the second is labeled observed, and they are never blended into a single image that would blur which is which.

7 · Rendering and output

From Blender to the browser.

Every representation leaves Blender as a glTF binary, the GLB format, exported with the Y-up option so the orientation matches the data, and with skinning and animation included for the bodies that carry weights and motion.9 REAL The standalone interactive viewer embedded above is vanilla three.js at version r160, which loads each GLB with the glTF loader, plays its animation with the animation mixer, and orbits the camera with the orbit controls, while the In World mode maps the generated panorama onto the scene environment as an equirectangular reflection. REAL

Inside the platform itself the three-dimensional studios use react-three-fiber, the React binding for three.js, to play the joint data directly as a drawn skeleton of joints and threads rather than to load these GLB files, so the GLB viewer and the studio playback are two distinct surfaces with two distinct jobs. REAL

8 · The eme mapping, as data architecture

Each primitive is a concrete data object.

The analytical primitives are not decoration on top of this build. Each one names a specific data object in the pipeline, and stating them plainly shows which parts are captured signal and which parts are produced around it.

The one primitive deliberately absent from this build is the facial one, the opseme, because the constraint that organizes the whole pipeline is identity removal, and the face is where identity lives. Its absence here is a design decision rather than an oversight, and it marks exactly where this build’s ethic and the facial primitive’s preservation ethic meet.

9 · Forthcoming work

What is described here and not yet built.

The strength of this account is that the line between what runs today and what is intended is drawn explicitly, claim by claim. The build is real and processed, and where it is not yet finished the method is described rather than asserted.