An open, evidence-based computational reference framework for biological form.

More than humans

The human is the first reference model, not the point of the project. What OMF is building is a framework for biological form in general: a way to say what a body's measurements are, how they covary, which combinations are anatomically possible, and what evidence backs each claim — in a form that holds for a sparrow as readily as for a person.

That generality is not a promise for later; it is a constraint on the design now. The anatomy vocabulary is bound to Uberon, a cross-species ontology, rather than a human-only one. The plausibility rules — hard bounds, a joint envelope, a validation gate — are stated over any organism's variables, not over human ones. And the schema is being stress-tested against a bird before it hardens, precisely because a mammal-shaped schema that has only ever seen mammals will quietly ossify around them.

Listing a species is not the same as having modelled it. Three of the four rows above are honest placeholders, and the page says so — which is itself the point of Unknown Is Preferable to Invented.

Surfaces

• Reference records

  Browse the variables, relationships, and citations that make up the reference models, with their honest review state.

  Open CC0 public domain

• Generator playground

  Dial or sample a profile → generate a rigged, shaded, validated human USD asset → view it → export it.

  Internal review password required

• USD viewer

  The browser USD viewer (OpenUSD on WASM). Point it at any published asset with ?file=/assets/<name>.usda.

  Open no Blender required

The boundaries

These are non-negotiable, and they are what make the reference models worth trusting. Several of them exist specifically to keep the framework from collapsing into a human character generator.

Evidence before opinion

Variables, relationships, and distributions trace to evidence. Every relationship carries its provenance: source, evidence strength, confidence, the population studied, review state, and version.

Unknown is preferable to invented

Where no reliable relationship is known, we mark it unknown rather than guess. A gap honestly marked is an asset; a plausible-sounding fabricated dependency is a liability that poisons every generator downstream of it.

Possible organisms only — no impossible forms

For every kingdom and phylum, not just humans. Enforced in three layers: hard per-variable bounds, a joint plausibility envelope over the covariance between variables (so combinations stay realistic, not just individual values), and a validation ruleset that re-measures generated output and rejects what escapes. Plausibility is a pass/fail gate, not an aspiration.

Form traces to measured evidence — no artist base

The generated surface is part of the scientific claim, not decoration. The test is instrument, not author: a real body read by an instrument (scan, photogrammetry, medical imaging) is evidence; a surface authored by an artist or a generative model is a plausible prior, never the base.

Align to established science

Borrow the structures science already maintains rather than minting our own — cross-species anatomy ontologies (Uberon, FMA, PATO) and measurement standards (ISO 7250 for humans; for birds, the protocols the ornithologists already use). We mint new terms only where science genuinely has none. Borrowing structure is encouraged; borrowing an artist's surface is not.

Generators consume; they do not define

Mesh, rig, texture, and medical generators read reference models. None of them ever writes authority back. Meshes, rigs, and textures are outputs — the product is the structured, open, evidence-backed reference knowledge itself.

Continuous variation, discrete compatibility

Biological variation is continuous; runtime profiles, topology families, and rig profiles are discrete. Keep the continuous model pure and quantize only at the generator boundary.

The full statement, including versioning and provenance: Founding Principles · Source repository Private