Nishi Human Program

→ BEYOND-METAHUMAN PROGRAM (live, updated each build round): /world/metahuman — the capability ruler, the honest distance to MetaHuman, the procedural face and skull built inside-out one measured unit at a time, with every rejected attempt published beside the shipped one.

Where we are, measured — and where we go, in the order the measurements demand. Every number here came from our own sovereign ruler; every failure is listed beside every win.

1. Where we are — measured, not asserted

The generator emits a human body from an anthropometric canon: 16 parts, ~341k triangles, three nested anatomical layers (bone core → muscle envelope → skin), digits produced by a parameterised rule, and the canon read from data rather than typed into code. It is measured against a real cadaver reference by nx_bodybench, which now runs five independent judges.

JudgeWhat it seesScore
front silhouetteoutline, face-on823
side silhouettedepth profile813
normal field (shape)surface orientation agreement923
detail (front + back)surface busyness vs the reference380
three-quarter detaillateral surface: flank, lats, arm/torso join255
head regionthe face alone227

Honest headline = the minimum = 255. Against a photograph rather than a cadaver, perceptual quality is roughly 175 — clay. The face is measurably our worst region.

The four gaps, ranked by evidence: soft-tissue form (the shapes that create up- and down-facing surface) · surface structure at the frequency a detail judge and an eye both see · face as true feature geometry rather than radius modulation · light transport, now started but starved.

2. What we learned — the laws that constrain the plan

Measured transfer moves proportion and cannot move detail without registration. Five attempts agreed: section priors, bin resolution, sampling method, fitted canon, high-frequency residual. The reference's anatomy at a coordinate is not our anatomy at that coordinate. This is why landmarks come first.
Light reveals geometry. Environment lighting on a body whose surface is all vertical has nothing to differentiate. The lighting gap and the detail gap were the same gap.
A metric is necessary, never sufficient. Every judge blind to something different has cost us the number and made it true: 837 to 313 when detail was added, 380 to 255 when the three-quarter view was. Hold several, publish the minimum.
Instruments before features. A region-scoped judge revealed that four facial bases moved the face by exactly nothing — after they had already been credited with a gain.
Read the source, not the summary. Twice in two sessions a primary source reversed a summary: a licence that was actually share-alike, and a "procedural" system that is actually diffusion-based on synthetic identities.

3. Where we go

Phase 1 — Instruments (unblocks everything else)

  1. nx_landmark — anatomical correspondence points (acromion, iliac crest, trochanter, patella, jaw apex, brow ridge, pupillary distance). Unblocks three sessions of measured-transfer results currently banked as failures.
  2. nx_campose — solve camera focal length and distance, replacing height-normalised framing so measurement is not warped by projection.
  3. Head oracle — the face currently has no reference at all. GNM is Apache-2.0; verify the data licence separately from the repository licence.
  4. Two-ruler split — anatomical realism against the cadaver, aesthetic against ratio targets. Stylised proportions should score worse on realism; without the split every style win reads as a regression.

Phase 2 — Form and identity

  1. Correlated manifold sampler — human variation is correlated and low-dimensional, so independent sliders produce implausible people. Source the structure from measured populations, never from a generative prior.
  2. Aesthetic scorer — canonical ratios, symmetry and feature harmony as an explicit measured permil, so beauty becomes improvable and red-teamable rather than hand-tuned.
  3. True facial geometry — a brow, orbit and nose are shapes, not a radius nudged by nine percent. Scored on the head judge and the head oracle.

Phase 3 — Surface and light

  1. Anatomically shaped relief bases, tuned against the three-quarter and region judges rather than a body average.
  2. Screen-space occlusion, soft area shadows, depth of field — deferred deliberately until geometry gives light something to reveal.

Phase 4 — Population

  1. Village generator: demographic structure, shared latents for family resemblance, generate-and-score so every inhabitant is plausible and distinct.
  2. Presets as points on the aesthetic ruler, never as edits to the realism baseline. Post-generation sliders are a later fine-control layer, not part of the generator.

Phase 5 — Past MetaHuman

Not by out-polishing Epic's artists. MetaHuman is a closed asset ecosystem with a DNA file binding morphs to anatomy. Our axis is the one nobody else has: a body correct from the inside out — the nested bone, muscle and skin stack where the outer form emerges from the layers beneath, extended with real muscles, a vascular tree visible through translucent skin, and organs. Infinigen's creatures are single-surface; SMPL is a single skin shell; MetaHuman is a shell with a rig. None of them can show what is underneath, because there is nothing underneath.

4. Posture toward outside work

Third-party systems are oracles and capability bars, never the product — the same posture that has served us with CUDA. Take the mathematics, which is published and free to learn from; take no assets, no weights, no runtime. A generated human that comes out of someone else's checkpoint is neither ours nor reusable, which is the whole point of building it.

Standing boundary: photographic references are a quality bar. We build look classes parametrically; we do not replicate the likeness of a named living person, which needs their consent.