nishi code wiki / research / digital twinning
Digital twinning of human anatomy: exact-person capture to surgical precision
Published · COMPLETES the in-progress brief opened 2026-07-31 · lineage: forks off cloth & soft tissue (SSM bars, damage bounds) and feeds the marble-doctrine pointing machine · goal: a data layer that describes a person EXACTLY — to duplication, to surgical use.
Freshness declared. Web-search budget exhausted this session: claims verified to the author's May-2026 horizon + the sovereign corpus; the May→July 2026 tail is UNVERIFIED and flagged where it most likely moved (generative face tooling, Gaussian avatar products). Re-verify +
nx_page_ingest archive pass is the declared next act. Every number below carries its measurement context.The finding that frames everything: exact duplication is a LADDER OF INSTRUMENTS, and each rung is already commercial or published. Clinical stereophotogrammetry holds sub-millimeter facial geometry TODAY and is already the standard of care in cosmetic-surgery planning; parametric shape models make identity a FITTABLE VECTOR; light-stage capture resolves pore-scale mesostructure. Nobody disputes the rungs — the moat is an INTEGRATED, sovereign data layer that carries all of them per person in one recombinable container.
1. Clinical capture — the surgical bar (SHIPPING)
| System | Class | Accuracy (published context) | Surgical use |
|---|---|---|---|
| Canfield VECTRA (M3/H2 face, WB360 whole-body) | stereophotogrammetry | face geometric error commonly validated ~0.2–0.8 mm (study-dependent; vendor claims finer) | THE cosmetic standard: breast/face/rhinoplasty simulation modules; pre/post outcome libraries |
| 3dMD (face/body, 4D dynamic) | stereophotogrammetry | ~0.2 mm RMS on craniofacial validation — the academic reference instrument | craniofacial surgery planning; 4D = expression dynamics captured, not posed |
| Artec Space Spider / Leo class | structured light | 0.05–0.1 mm point accuracy (small field) | prosthetics, part-scale anatomy |
| Crisalix | photo/scan → cloud 3D sim | consumer-photo capture; validity studies show simulation-vs-outcome deviations at the several-percent-volume level (breast) | patient-communication simulation (breast/face); VR consult |
| CT / MRI | volumetric | sub-mm voxel | the INTERNAL twin: implants, bone — the only rung that sees under the skin |
Verdict: sub-millimeter external geometry is a solved, purchasable capability. The surgical-precision claim for OUR tooling means matching this bar with measured error, not vendor language.
2. Parametric identity — the person as a fittable vector (RESEARCH → shipping edges)
| Line | Status | The number that matters |
|---|---|---|
| Head/face 3DMMs: FLAME; image→identity fitters (DECA/MICA lineage) | RESEARCH, mature | NoW benchmark median reconstruction ~1 mm-class from a single image; REALY region-wise ~1 mm — single-photo identity is a MILLIMETER problem now |
| Body models: SMPL-X / STAR lineage over CAESAR-class corpora | RESEARCH, mature | scan-fit ~2–3 mm surface; image-fit ~5–10 mm (pose-dependent, non-rigid) |
| Breast SSM (Weiherer et al., The Visual Computer, open, CC-BY) | RESEARCH, open | 0.17 mm generalisation / 2.8 mm specificity — already banked as our oracle; a LINEAR model holds sub-mm on the exact tissue cosmetic surgery cares about |
| Neural parametric heads (NPHM class), Gaussian codec avatars (Meta relightable line) | RESEARCH, 2024–25 | sub-mm surface detail + relightable appearance; heavy capture rigs; the fidelity ceiling demonstrations. ⚠ most likely to have moved May→July 2026 |
| Dense landmarks (700+-point face sets, MediaPipe-468 class in every browser) | SHIPPING | the POINTING MACHINE's off-the-shelf detector leg |
| Light-stage mesostructure (polarized gradient illumination) | RESEARCH, standard in film | ~0.1 mm pore/wrinkle displacement — the surface-pass ceiling |
3. What EXACT DUPLICATION requires the data layer to hold (the NXA answer)
| layer | content | precision bar | NXA section |
|---|---|---|---|
| proportion vector | skeletal/anthropometric parameters | ±1–2 mm segment lengths | PARM (new) |
| surface geometry | identity mesh + per-vertex displacement vs the parametric base | 0.2–0.5 mm (clinical bar) | FACE + DISP (new) |
| mesostructure | pore/wrinkle displacement + normal maps | ~0.1 mm | TEXM |
| appearance | albedo/spec/rough per region; capillary zonation | ΔE < 2 per region (colorimetric, not eyeballed) | TEXM |
| expression space | personal blendshape basis (HER smile, not generic) | NoW-class ~1 mm on expressed states | MORF |
| dynamics | per-region soft-tissue k/c/travel — HER tissue response | frequency ±0.2 Hz, amplitude ±10% vs subject video | DYNA (live) |
| internal (surgical rung) | bone/implant volumes where consented imaging exists | CT voxel | VOLM (future) |
The moat restated: every competitor holds SOME layers (Vectra: surface+appearance; FLAME: vector+expression; light stage: mesostructure) — none ships ONE portable, self-verifying, engine-agnostic container carrying all seven with in-asset physics and integer-deterministic replay. That container is NXA's endgame — and it is precisely what licensed acting + surgical consult both need.
4. Our climb, in the marble order (each pass gated by the pointing machine)
| pass | act | tolerance gate (falsifiable) |
|---|---|---|
| 0 · pointing machine v0 | dense landmark detector over reference frames AND our portraits; per-point error table in mm at subject scale | instrument agrees with itself ±0.5 mm across re-runs; detector error known vs 50 hand-marked points (the validation-set discipline) |
| 1 · MASS | skull/head + body proportion fit against the canon + subject | proportion vector inside ±2 mm; NO surface work admitted before this passes |
| 2 · SECONDARY | feature volumes: orbit, nose (rhinoplasty constants), lips, breast SSM fit | per-landmark ≤1 mm on the fitted regions (NoW-class) |
| 3 · RELIEF | displacement detail vs scan/photo stereo | ≤0.5 mm surface RMS (clinical bar) |
| 4 · SURFACE | mesostructure + colorimetric appearance | ΔE<2 per region; pore-scale where capture supports it |
5. The corpus question: can photos/videos/GIFs reach the surgical bar? (operator 2026-08-03)
Answer: a three-tier ladder — and the binding constraint is never pixel count. It is (a) SCALE ANCHORING and (b) per-region honesty about measured-vs-inferred. A clinical rig is just photogrammetry with controlled calibration, sync and light; casual imagery gives up those controls and the fit must account for what was lost.
| tier | source material | honestly achievable | why / what binds |
|---|---|---|---|
| T1 class | GIFs, 360p video (the bulk of the corpus) | cm-scale proportions + dynamics bands (we already run this: the 2.5 Hz oracle, class manifolds) | compression destroys the geometric signal (our own alias findings); single views are depth-ill-posed. Surgical: NO. |
| T2 consult-grade twin | HD video of a cooperative subject, multi-angle, minutes of footage | ~1–3 mm surface, ~1 mm per-landmark on well-seen regions via multi-view + neural refinement (SfM/MVS COLMAP-class SHIPPING; monocular-avatar research line 2023–25) with non-rigid handling | this MATCHES the photo-class tools surgeons already use for patient communication (Crisalix class). Binds on: scale anchor (interpupillary prior ±2–3% unless a known-size object appears), subject deformation between frames, unknown lighting biasing shape-from-shading. |
| T3 sub-mm on best material | professional photo SETS (studio stills, RAW-class resolution, many angles — the high-end galleries) | sub-mm approachable on RIGID, well-textured regions; the strongest non-scanner source in existence | still needs a scale anchor for absolute units; the verified 0.2 mm CLINICAL claim stays scanner territory until a validation study says otherwise — we do not claim it from imagery. |
The design consequence that makes the corpus usable for surgical conversation anyway: the twin fit must output PER-REGION CONFIDENCE — “nasal dorsum: measured, 0.9 mm from 214 frames; retro-auricular: prior-inferred, ±3 mm” — so every downstream use knows which millimetres were SEEN and which were FILLED by the shape prior. Priors keep improving (SSM→neural parametric lines), which moves T2 toward T3 over time without any new capture; the error-bar layer is what keeps that honest. This is also the exact leg the pointing machine v0 builds first.
UNVERIFIED / declared gaps
- May→July 2026 tail unverified (search budget). Most volatile: generative face-forming tools (the operator-referenced Google-class tooling), Gaussian avatar productization, any new NoW/REALY leaders. Re-verify pass queued with citation archiving.
- Accuracy figures above are study-context-dependent: stereophotogrammetry validation numbers vary with landmark protocol and surface region; the brief reports the commonly-validated bands, not best-case vendor claims. Each figure gets a pinned citation in the archive pass.
- Surgical-simulation OUTCOME validity (does the predicted breast match the post-op breast) is measured in the literature at the several-percent-volume level — good enough for consult, NOT yet a guarantee; our error-bar framing (reachability as a number) matches what the field can honestly support.
- Consent/licensing layer is structural, not legal advice: capture-to-duplication of a real person proceeds ONLY under license; the throwaway class-calibration path never stores an identity vector.