Mosaic portrait of Socrates assembled from code symbols and typographic fragments

Tsevis Studio

Algorithmic Mosaic

Up close, the viewer reads the parts. From a distance, the whole. Both experiences are designed.Tsevis Studio

Opening image: Socrates, algorithmic mosaic for Communications of the ACM, Tsevis Studio.

An algorithmic mosaic is an image built from thousands of smaller elements, placed by a computational method so that the whole can be read from a distance and every part keeps its own meaning up close. The elements can be photographs, symbols or typographic characters. The method decides which element goes where, so the result can be repeated, measured and refined.

Tsevis Studio, led by Charis Tsevis and Vira Konstanta, is the contemporary practitioner of Computer-Assisted Mosaic Tessellation, the tradition originated by Ken Knowlton at Bell Labs in 1966. We build algorithmic mosaics at institutional scale for sovereign commissions, global brands and permanent architectural installations, using proprietary engines developed over thirty years of studio practice. The same method serves a magazine cover, a campaign, a building entrance and a hand-finished work on paper.

What makes a mosaic algorithmic

A mosaic is algorithmic when a rule decides where each fragment goes. The rule can be a grid, a matching process, a map of where the eye looks, or all three. The fragments can be images, symbols or typographic characters. The result is not a filter laid over a photograph. A filter changes the pixels of one picture; an algorithmic mosaic replaces them with chosen elements, each of which can be read, and the choice of elements is as much a part of the work as the image they form. It is a structure with two designed experiences: the parts up close, the whole from afar. Because the rule is explicit, it can be changed: a different grid, a different matching method or a different map of importance gives a different mosaic from the same image.

The four layers

Every mosaic, ancient or digital, works on four layers at once. We design all four on every commission.

  1. The main image. What the viewer perceives at a distance: the face, figure or message.
  2. The tesserae. The individual units and what they carry. In our work they are images, symbols and characters chosen as cultural memory, not as texture.
  3. The constructing mechanism. The grid, algorithm or logic of assembly. Much of the thinking happens here, and it is often invisible in the finished work.
  4. The semantic layer. The meaning that binds the other three: why these fragments, for this image, for this client, at this moment.
Mosaic portrait of Steve Jobs assembled from images of Apple products
2008: Steve Jobs for Fortune, built from Apple products. See the work
Mosaic portrait of Barack Obama assembled from numbers, statistics and financial symbols
2016: Barack Obama, built from numbers, statistics and financial symbols. See the work
Portrait of a smiling woman with glasses, her face formed entirely from typographic characters
Mary Higgins, typographic portrait for Emerson College, 2024. The face is built entirely from typographic characters. See the work

The full argument is in The Tradition We Work In, section IV.

Algorithmic Gestalt and the Macro-Micro Narrative

Algorithmic Gestalt is our term for applying perceptual theory (emergence, figure and ground, proximity) to the computational design of mosaic systems. Its practical rule is the Macro-Micro Narrative: every scale of a work must carry and amplify its central message. See section X of the Tradition essay.

The same portrait, rebuilt on two different grids, shows the idea at work: the whole is read from afar, and each structure changes what the eye finds first.

A smiling young man with curly hair, shown as a painted portrait on the left and as a square-tile mosaic with black and white lattice lines on the right
One portrait, a square grid. The left half is the source portrait; the right half is the same image rebuilt from flat-colour tiles on a square lattice, with a black and white structure between them. Seen at reading distance, the eye closes the gaps and finds the face and the shape of the hair against the sky. Seen up close, the lattice itself becomes the subject. The grid is the constructing mechanism, the third of the four layers described above.
The same portrait as a mosaic built from hexagonal cells, each a cube-like glyph in three tones, with the painted source portrait on the left
The same portrait, a different structure. Here the right half is built on a hexagonal lattice, and every cell is a cube-like glyph drawn with three tonal faces, so light, middle and dark values are built inside the tile and not only between tiles. The eye, the nose and the smile are still found from a distance, although no single tile contains a feature: figure separates from ground, and neighbouring cells of similar tone group into forms. Changing the grid changes what the viewer notices first, which is why we treat it as a design decision and not as a filter.

A long tradition, not a trend

Mosaic is among the oldest ways humans have built images, from the clay-cone walls of Uruk and the pebble floors of Pella to Byzantine Ravenna, Futurism, typewriter art and Knowlton’s computer mosaics. We work inside that line, and we say so. Read the tradition.

How we build it

We build our own instruments, because existing software did not address the problem at the depth the work needs. They answer three questions: which fragment belongs where, where the eye goes, and which grids and pattern traditions give a mosaic its structure. We measure, gate and document the method. The full account, with the science we credit, is on the research page.

Three of those instruments are shown below as the studio’s own settings windows: how detail is protected in a face, how importance is measured, and how a single tile is chosen.

Settings window for priority-aware dithering, with a black and white priority map of a portrait at the centre and sliders for edge, blue-noise and alpha control on either side
Where detail is protected. Dithering decides how colour error is passed on to neighbouring tiles. The Atkinson kernel, devised by Bill Atkinson for the original Macintosh, passes on three quarters of the error and gives a stippled, graphic surface. The centre panel is a priority map of a portrait: white marks the eyes, nose and lips, where dithering is reduced so the face stays exact (25 percent in the setting shown), and black marks hair and background, where full dithering (100 percent) lets the surface breathe. The right-hand column adds edge-aware control, blue-noise variation that behaves like film grain, and alpha-aware matching for transparent tiles. Settings window, Tsevis Studio.
Importance analysis window showing twelve colour-coded maps of a smiling man's portrait and a large combined map, with checkboxes and sliders for each method
Where the eye goes. Before any tile is placed, the software measures importance in many ways at once: edge detection, saliency, structure tensor, colour variance, texture, frequency, contrast, face awareness and an AI-enhanced detector, each shown as its own panel. The weighted maps merge into the combined map on the right, which rates every region of the image and feeds the priority system below it (an influence control and a response curve). Presets (Portrait, Landscape, Balanced, Subject AI, Poster) set the starting weights, and plain-text instructions such as “keep detail: eyes, lips, teeth, hair” steer the result. This is the practical side of the importance and Gestalt questions on the research page. Settings window, Tsevis Studio.
Settings window for tile matching with a Portrait preset, luminance gate, dark region protection, performance options and a neural reranker section
How a tile is chosen. Each tile is matched by its internal 8 by 8 layout of colour and luminance, not by a single average colour. The Portrait preset keeps a tight luminance gate (about 11) for a crisp likeness; loosening it towards 22 admits more tile variety. Dark-region protection keeps tiles that are too bright out of shadow. An optional second stage re-ranks the colour-plausible candidates with a self-supervised vision model (DINOv3), weighted here at 30 percent, so that a tile is chosen for how it looks as well as for its colour. The stage is resource-intensive and runs on studio machines. Settings window, Tsevis Studio.

Where it is used

One method, many rooms: architecture and public art, corporate and institutional commissions, gifts and editions, advertising, editorial and science communication, sport and culture, healthcare. Clients include Nike, Bulgari, Delta Air Lines, the United States Olympic Committee and the Royal Mint. See the applications.

When the mosaic leaves the screen

Singularis is the work Vira Konstanta and Charis Tsevis make by hand: algorithmic mosaics rebuilt in relief paper, gold leaf and archival inks, each unrepeatable by design. See Singularis.

Who we are, and how to start

We are based in Europe and work with clients worldwide. About the studio · Start a brief.

Further reading on the blog