Skip to main content
Book a call

AI Portrait Generation Technology

The science behind the art

The science
behind the art

We post-train open image models on our own data, build the software that runs our kiosks and Studio stations, and operate the GPU servers our models run on.

Black-and-white gallery wall with blue pixel pattern

Built in Rust

Built
in Rust

Two renderers from the Rust image library our team maintains, shown on Vermeer's Girl with a Pearl Earring instead of a visitor's portrait. Its pixel-art step is a port of PixelOE by Shih-Ying Yeh, released under Apache-2.0.

Brick portraits

At The Art of the Brick, each generated portrait is rebuilt from 8,175 bricks. Eyes and other thin details are thickened so they survive the downscale, and every brick takes one of 16 colours, never dithered. Moved to Rust in February 2026, the whole step takes under half a second on a desktop CPU, against 1.25–1.79 s before.

  • Rust
  • Outline expansion
  • In production
Learn more

Drone-light portraits

A prototype souvenir that pictures a portrait as a drone formation; it plans no flights. A depth model estimates the shape of the face, and the renderer spreads a hexagonal lattice over it in perspective, so nearer points grow and brighten. The render takes about half a second on a desktop CPU, the full pipeline a median of 16–21 s.

  • Prototype
  • Rust
  • Depth map
Learn more

Technical capabilities

Technical
Capabilities

The building blocks behind every portrait and every live deployment.

Recognisable faces

We compose the group portrait first, then refine each face using that person's photo to preserve their likeness.

Portraits in seconds

From confirming who is in the photo to the first portrait on screen: a median of under 5 seconds on production kiosks, measured June–September 2026.

Custom portrait styles

We tailor portrait styles to the venue. Our team tests them across a range of faces and reviews the sample portraits one by one.

Print and digital output

Portraits are prepared for printing and digital delivery, with print frames customised to the venue.

Ready for busy days

Portraits run on GPUs we operate. When demand grows, the platform automatically adds cloud GPUs running the same software and releases them when the queue clears.

Monitored remotely

We monitor kiosk connections, GPU health and print jobs remotely, with alerts when a print needs attention. Software updates run between visitor sessions.

Measured in production

Measured in
production

Six figures from our production systems, each with exactly what it counts.

Portraits generated
4M+
Portraits generated for visitors in the twelve months to September 2026, at our installations and through our API. Counts images, not people.
Sessions with a portrait
679K+
Visitor sessions that produced at least one portrait in the same twelve months. A session is one interaction, not a unique person.
Busiest day
31K+
Portraits generated on the busiest day of those twelve months, 3 January 2026: about three times a typical day.
First portrait
<5S
Median wait on production kiosks from the visitor confirming who is in the photo to the first portrait on screen, in every month since June 2026.
Production styles
195
Catalog styles switched on for venues as of September 2026. New styles start switched off, and only an administrator can switch one on.
Interface languages
10
Languages the kiosk interface ships in, from Dutch and Greek to Chinese. Each venue chooses which ones its visitors can switch between.