Geometry you can question.
Somaiva is an experimental workspace for reconstructing and comparing external facial surfaces from photographs. Its purpose is to make geometry inspectable, repeatable and honest about uncertainty.
How a reconstruction works
Your browser decodes each photograph, removes metadata and runs the locally served MediaPipe Face Landmarker. It checks face count, image quality and view coverage before you reconstruct.
After you explicitly choose Reconstruct, only landmark coordinates and quality metadata go to the server. The geometry engine aligns and robustly fuses the observations, builds a consistent head surface, and calculates documented measurements.
The face is estimated from observed landmarks. The unobserved cranium is generic geometry; ears are not reconstructed. A temporary, optional single-frontal photograph can be projected onto the observed face. It is an approximate display texture, not multi-view blending; generic cranium regions remain neutral and texture never changes measurements.
What the numbers mean
Measurements describe external soft-tissue and surface geometry. Scale is normalized so facial height equals 1. There is no verified metric calibration, and values are not millimetres.
Depth is partly inferred by the landmark model. Unknown lenses, camera distance, makeup, retouching, age differences, expression and occlusion can change the result. Coverage scores and confidence proxies do not establish surgical accuracy.
MediaPipe supplies no calibrated confidence for every landmark. The displayed proxy combines conservative quality and consistency signals. Unsupported measurements are withheld.
What it cannot tell you
Photographs do not reveal skull structure. This tool cannot determine which procedures you need, predict clinical outcomes, identify a person, or establish attractiveness or gender.
Gender-perception model not yet available. No masculinity, femininity or “passing” rules are hidden in these measurements.
Geometry sandbox sliders only deform a rendered surface. They are not medical recommendations or a surgical simulator.
Private by design
Raw photographs never leave your browser. Canvas re-encoding strips EXIF, including GPS, before analysis. Photo previews live only in memory. After reconstruction, one cleaned frontal preview per active set may remain for the optional texture toggle. All other previews are removed. Clear texture, Delete session, navigation, refresh or a new reconstruction releases retained previews. Reopened sessions have no photo texture. GLB and JSON exports contain geometry without photographs.
Landmarks are sensitive derived facial data. They are sent to Somaiva only after your explicit reconstruction action, processed in memory and not written to a server database. Request logging excludes photo bytes and landmark arrays. We do not train on uploads or use analytics.
Meshes, landmarks, measurements, names and dates are saved in your browser's IndexedDB. Delete Session removes them from that browser. Downloaded files are under your control.
Suitable photographs
Use 3–30 JPEG, PNG, WebP or AVIF photographs of the same consenting adult. Include frontal, left and right three-quarter views; profiles can improve coverage when detectable. Use a neutral expression, consistent appearance and unobstructed features.
HEIC is not supported in this version; export it as JPEG. Each image is limited to 20 MB and 32 megapixels. Images are resized to at most 1600 pixels on the longest side. Duplicate photographs are excluded. Repeated similar poses add little coverage; additional images do not guarantee greater accuracy.
Animals and objects usually produce no usable face. Cartoon and synthetic faces are unsupported and are not reliably distinguished automatically. Sunglasses, severe occlusion and appearance inconsistencies need manual review.
Built for evidence
Each measurement exposes its landmark definition in the geometry lab. Comparison uses stable-landmark alignment and optional global scale normalization; it does not force entire surfaces into agreement.
Synthetic benchmarks test deterministic geometry and regression, but do not prove identity fidelity or clinical accuracy. In the procedural-head image benchmark, MediaPipe detected all five rendered views, yet chin projection error reached 78%, nose projection 29% and jaw width 11%. Detection success is not measurement validity. Real multi-view ground truth and repeat captures are needed to validate the reconstruction approach.
Reconstruction, appearance analysis and future human review are separate interfaces. Payments or referral relationships cannot alter geometric measurements.
Explore the surface. Inspect the assumptions.
Start with your photographs, or open a synthetic specimen in the geometry lab.