This person does not exist.
Neither does anyone else on this page. An AI invents every face from random static, with the features you choose, on your own graphics chip, right here in the browser.
How it works
From static to a face in 25 small steps
- Random staticEvery face starts as pure noise, like an untuned TV.
- Cleaning upThe AI removes a little of the noise, 25 times. Scroll to watch each draft.
- A new faceSomeone who has never existed. Not a copy: 0 of 10,000 checked faces matched a real photo.
- Your choicesAsk for glasses, a smile, grey hair, bangs. A second AI checks each one.
01Make faces
Invent someone new
Choose features each face must have or must not have. A second AI reads every face before you see it, sends back any that missed, and marks what each one has.
02Instant or best
One step, or twenty-five?
The same starting static, drawn four ways. Instant mode is a second AI trained to jump straight to the answer. Best quality takes more clean-up steps: slower, but closer to a real photo. With live mode on, the race runs on your own graphics chip.
Realism gap: how far the faces are from real photos, measured on 50,000 faces (FID, lower is better).
The faster your GPU, the faster the faces
Instant faces a minute, each measured through this page. Every face is made on the visitor's own device; nothing is uploaded.
03Is it AI-made?
A signature you can't see
Every face this site makes carries an invisible watermark. It survives resizing and compression. Drop any picture here, or try the two examples: the check runs on your device, and the picture never leaves it.
Examples
Waiting for a picture
The answer appears here.
04For engineers
What's under the hood
A 183-million-parameter diffusion transformer that draws in the feature space of a frozen vision model, distilled into a one-step generator. Every number below is measured.
Realism gap of every model, measured
FID (clean) against FFHQ-256 on 50,000 faces each, log scale, lower is better. Orange: what this site runs. Green: the best.