Pratima turns a photograph into a fully-rigged 3D human, running in real time on a standalone headset — in about fifteen minutes. Not generated art. Measured geometry, checked against spec, and refused when it doesn't meet it.
The natural assumption is that a neural network does the whole thing — a photo goes in, a character falls out. That's not how Pratima works.
Exactly one of six stages invents anything: a generative model that turns your photograph into a rough 3D head. Everything after it is arithmetic — fitting a template, projecting pixels, deciding per point whether a surface faced the camera.
Run the same photograph through twice and the output is byte-identical, down to the last bit. That's what lets the engine debug itself, gate its own output, and refuse a face it can't do well — instead of quietly shipping one that looks wrong.
A cause we can find and name — not a model we can only retrain and hope. When one character's jaw looks wrong, we can point to exactly why.
~40 automatic checks compare the result against spec. Any one can refuse the character — the engine stops rather than ships something subtly wrong.
Improving the engine never silently changes a character you've already shipped. What you integrated is what stays integrated.
We never generate a body. Your photograph supplies identity — expressed as numbers that deform a template we authored, with a fixed topology and a fixed 26-bone skeleton.
Author a motion once and it plays correctly on every character the engine will ever make — which is why we consume NVIDIA Kimodo's text-to-motion output instead of hand-animating each one. A free-form generated mesh can't do this; ours can, because the skeleton never changes.
Built and running today, and what's still in front of us — plainly.
“Photo-to-avatar platforms have a habit of getting acquired and shut down overnight — developers lose their pipeline with days of notice. Pratima runs as software you operate yourself, on your own infrastructure, so that can't happen to you.”