METHOD & PROVENANCE

A study in editable form.

This prototype builds an adult clay mannequin from original mathematical surfaces. A head built from authored cross-sections has independent skull, jaw, chin and forehead controls. Orbital apertures expose spherical eye globes wrapped by upper and lower eyelid rims. Authored fields shape the nasal root, bridge, tip, alar wings and nostril recesses. Separate three-dimensional upper and lower lip surfaces include a Cupid’s bow and mouth corners. Auricle relief includes a helix, concha depression, antihelix, tragus and lobule. A shorter, broader neck blends into a rounded shoulder mantle. These are authored approximations, not validated anatomical reconstruction. There are no scans, learned weights, downloaded meshes, external likeness datasets or model-library assets.

What a measurement means

The reference is Synthetic 02: a fixed set of deliberately authored parameters, unrelated to a real person. Front, 45° and side orthographic head silhouettes are rasterized at 96 × 128 pixels in a fixed 260 × 300 mm measurement window. Silhouette error is 1 minus intersection-over-union, averaged across the three views. Landmark error is the root mean square of projected two-dimensional distances across eight authored surface landmarks and three views, in millimetres. The head, eyes, lips and ears contribute to the silhouette; the neck and shoulder mantle are excluded. Measurements in v0.2 use Synthetic 02 and must not be treated as an error reduction from the v0.1 target.

Each committed parameter edit or sculpt dab records baseline and candidate states, measured errors, geometry complexity and sampled rendering cost. Positive error reduction means closer to this particular synthetic target. A finite raster can miss small improvements; landmarks cover only a few points. Neither metric measures beauty, anatomical truth, identity or exact likeness. No aggregate score is computed, and no future gain is predicted.

Different forms, one material

Six clearly labeled fictional adult presets provide oval, angular, round, tapered, contoured and controlled-asymmetry starting points. They vary skull and jaw breadth, chin, cheeks, brow, orbital opening and lid hood, bridge/profile/tip, lips and authored adult contour cues. They all use the same clay and lighting; no preset encodes ethnicity, identity, attractiveness or an inferred age. Every preset remains editable. Adult contour cues are design parameters, not biological age estimates.

Native schema valkrai-face/0.2.0 can read v0.1 dimensions by filling newly added controls with authored defaults. It does not claim to reproduce the old generator’s exact geometry; that generator is retained in the release rollback. Browser JavaScript is inspectable even when canonical source is held locally or in a private repository.

Coarse to fine, with a way back

Structure, Features and Detail keep broad proportions accessible before local edits. Each sculpt dab on the front side of the head has a separate radius in millimetres and signed displacement amplitude in millimetres, applied through an original Gaussian field. Halving radius and halving amplitude are separate controls. If amplitudes follow a, a/2, a/4, …, their sum is bounded by 2a. This bounds accumulated displacement magnitude under unit directions, but does not guarantee target reachability, improvement, convergence to a useful shape or preserved surface quality.

The experimental adaptive button uses the last measured silhouette gain. Above 0.0005 (0.05 percentage points), it keeps the current scale. Otherwise it halves amplitude and multiplies radius by 0.75, subject to minimums. It does not undo a worse edit or predict a better one. The history tools let you reject that edit, and Structure remains available to revisit coarse form. A stronger successor would evaluate several candidate directions and scales against all views, then accept only an explicitly improving candidate. Radius and amplitude must remain independent.

Rendering cost and limits

Rendering measurements use this browser’s performance clock, requestAnimationFrame and gl.finish. They include CPU geometry reconstruction, buffer upload and completed drawing, with 12 warmup frames. They exclude interface updates and geometric diagnostics. gl.finish is a synchronization point, not a GPU-only timer. Per-edit draw samples are noisy. Resizing a desktop browser to a mobile viewport is not physical iPhone validation.

This is a stylized analytic model, not a production character system. The eye/lid rims, lips and ears remain overlapping procedural parts. Nostril recesses are sculpted surface depressions rather than full nasal cavities. Ear attachments and facial tissue transitions remain simplified. Extreme parameter combinations can show feature intersections; exact triangle self-intersections are not checked. It has no muscular anatomy, hair simulation, skin tissue model, facial performance capture or speech synthesis. OBJ exports a static surface; native JSON preserves editable dimensions and dabs. A future glTF adapter would require documented topology, rig and expression mappings.

Historical context

Simple ratios can be useful starting constraints, but are not universal facial or female anatomy. The Royal Collection Trust describes how Leonardo’s measurement studies diverged from his search for simple ideal proportions. This lab does not reproduce his drawings or assume a single canonical face. Leonardo’s Study of Anatomy — Royal Collection Trust.

Back to the studio ↗