How you measure a rig you do not have
A reference “look” ships no body geometry. The base figure is licensed content that lives inside the host application, so the package carries only deltas against it. That sounds like a dead end for comparison; it is not.
Two things leak the base figure's specification:
- The binary morph format is vertex-indexed. It is a 4-byte header followed by 16-byte records of
[int32 vertexIndex][3 × float32 delta]. We confirmed that layout by arithmetic rather than assumption: a 202,020-byte file yields(202020 − 4) / 16 = 12626records, exactly matching thenumDeltasthe metadata declares. The largest vertex index observed is 21,530, so the base figure carries at least 21,531 vertices. - Morphs adjust bones, and name them. Each morph's metadata carries formulas targeting skeleton bones by name (
BoneCenterX/Y/Z,OrientationX/Y/Z). A morph cannot target a bone that does not exist in the skeleton it was authored against, so the formula targets enumerate the rig.
Three independent packages yielded 80, 80 and 83 unique bone targets. That is a floor, not a census — only bones some morph happens to adjust appear at all.
The comparison
| Rig feature | Reference figure | Ours today |
|---|---|---|
| Base mesh vertices | ≥ 21,531 (derived) | 14,164 (measured) |
| Joints | ≥ 80 named in morph formulas | 104 declared |
| Per-finger hand rig | Yes — thumb/index/mid/ring/pinky ×3 + 2 carpals per hand | — |
| Per-toe foot rig | Yes — big toe, 4 small toes, toe, foot | — |
| Jaw bones | Yes — upper and lower jaw | No |
| Eye bones | Yes — left and right | No — eyes are a separate sphere pass |
| Tongue | Yes — 7 bones (base, 01–05, tip) | No |
| Lateral chest pair | Yes — lPectoral / rPectoral | No — zero off-midline joints at chest height |
The finding that explains a workaround we already had
Our own asset-floor gate reported it earlier the same day: at chest height our skeleton has five joints, and all five are on the midline — the spine and sternum chain. The lateral pair is zero. Driving those midline joints swings the whole torso, which is precisely the defect that was observed and worked around months of iteration ago.
The remedy adopted then was sound and is now explained: soft tissue with no bone of its own has to be moved at the vertices it occupies. That is why our chest motion is a vertex-band effect rather than a joint rotation.
What the reference rig shows is that this was never a physics problem. Their figure has lPectoral and rPectoral as first-class bones, referenced 12 to 30 times across the morph sets we examined. We had correctly diagnosed “this rig has no breast bones” and correctly built around it — but we had filed it as a shading and simulation constraint, when it is a skeleton constraint with a known shape and a known fix. Our asset-floor gate already carries the tooth that will confirm it: a lateral-pair check that goes green only when one bone is declared left of the midline and one right of it.
Why the exceed path is generation, not import
The reference body mesh is licensed content and is not present in any package. We can measure that rig forever and never render it, which settles the strategy rather than limiting it: there is no version of this where we import our way to parity, so the only route is to generate a rig that surpasses it.
The next tier above the reference is the industry facial-rig standard, whose joint counts run several hundred and whose value is concentrated in the face — exactly our measured worst region. That is the target to research and exceed, and the honest sequence is: close the named skeletal gaps first (lateral chest pair, jaw, eyes), because a facial rig built on a skeleton missing its jaw has nothing to attach to.
UNVERIFIED — what this page does not establish
- The 80–83 bone figure is a lower bound, not their skeleton's size. It counts only bones that some morph adjusts. The true count is higher and we did not measure it.
- Not all packages expose it. Several looks ship morphs with no bone formulas at all, so their rig is invisible to this method. Our first sweep reported zero across the corpus because it searched file names rather than file contents — a detector is defined as much by where it looks as by what it matches, and the corrected sweep is what the numbers above come from.
- No side-by-side 3D render exists. We can render ours and cannot render theirs, so any visual pairing would compare our render against their specification. We publish the specification instead of staging a picture that implies more than we measured.
- Vertex count is not quality. A larger base mesh is not automatically a better one, and nothing here measures deformation quality, topology or edge flow — only counts and presence.
Sources
- Virt-A-Mate Hub, free resources — the 50-package corpus measured
- Belline Ultimate — Androinz (CC BY-SA) — morph formulas yielding 80 bone targets
- Mackenzie — PetaZwega (CC BY) — independent confirmation at 83 targets
- Character surface floor — the companion measurement, one layer out
Lineage — Forks off the character surface floor measurement of the same day, which established the corpus and the measure-only doctrine. Parent: character asset floor, 2026-08-03.