Every Tile Is a Basketball: A Study

Written by Charis Tsevis

A basketball player, rebuilt from basketballs. Every tile is a black-and-white ball, and the player is whatever the balls make together. This is a study, not a finished piece. The source image is by my friend and collaborator Paolo Guidobono. We have both worked for Lega Basket for more than ten years, and the study grows out of our neofuturistic work for them, Lega Basket Serie A: The New Futurists.

Black on white. The ball in the clip leaves a wake of bigger balls.

The idea

Each tile is an eight-panel basketball. Dark areas of the picture get balls with more black showing, light areas get more white. The principle is old: Harmon and Knowlton built pictures from symbols of different density at Bell Labs in the 1960s [1].

A basketball cannot be white, because its seams are black. The lightest ball I built is about a fifth black. So where the picture is lighter than that, there is no ball at all, and the white of the page does the work.

Thirty black-and-white basketball tiles in three lines of ten: the same ball in six panel colourings, each seen from five angles.
The tiles: one ball in six panel colourings, two to a line, five views of each. A seam is white only when both panels beside it are black.

Four looks

Four clips, four looks: black on white above, then white on black and two in colour.

White on black. The dark parts of the picture stay empty and the light parts become balls.
Colour. Each ball takes one colour, the average of the picture under it.
Colour, a second clip.
Four vertical mosaics of a basketball player made of basketball tiles: black on white, white on black, and two colour versions.
One moment of each of the four clips, frame 110: black on white, white on black, and two in colour.

Stills and details

Three black-and-white mosaic frames of a basketball player in motion, built from basketball tiles on white, with large balls at the left and right edges.
Frames 20, 110 and 160 of the first clip. Small balls make the player; the large ones sit at the edges.
Left, a basketball player in a light blue jersey with the number 3 dribbling against an orange and blue swirl. Right, the same moment as a colour mosaic of basketball tiles on white.
The source image by Paolo Guidobono at 1.5 seconds, and the colour mosaic of the third clip at the same moment.
Close-up of small dark basketball tiles forming a head and shoulders, with a few large blue and grey balls beside it.
Detail, enlarged. The head is made of the smallest balls, each with one colour.
Close-up of large basketball tiles in orange, pink and blue, getting smaller toward the upper right.
Detail, enlarged. Large balls at the edge of the picture, each with one colour.
Close-up of black-and-white basketball tiles packed densely into the shape of a head and torso on white, with larger balls at the upper right.
Detail, enlarged. Black on white: the figure is the small balls.
Close-up of white basketball tiles on a black ground, with the dark area of a face left empty.
Detail, enlarged. On black, the dark parts stay empty and the light parts become balls.

Seeing the ball

The computer vision is small and has one job: find the ball, and let the rest follow.

  • The ball. SAM 3 [2], prompted with the word “basketball”, on every second frame, interpolated in between. It found the ball in 106 of the 107 frames I asked it about.
  • The wake. Where the ball is now, and where it was in the last three quarters of a second, the balls around it grow and thicken.
  • The player. In the first stills I cut the figure out with BiRefNet [3] and kept the orange halo around it as a motion trail, its fade measured along and across the trail with scikit-image [5]. In the clips I dropped the matte: the size of the balls follows the edges of the picture, and the top 20% and bottom 15% of the frame keep only the small balls and the wake.
  • The white. Where the picture is lighter than the lightest ball, the circle stays empty and its tone error moves to its neighbours: error diffusion [4], run over circles instead of pixels.
  • The weakness. Tiles are so close in tone that a tiny change in the picture picks another ball. In an early test, a change of 0.3 grey levels in a still background altered 22% of the pixels from one frame to the next. Holding the circles still helped. It did not cure it, and the clips still twinkle a little.
Animation of a black-and-white basketball mosaic of a player in motion, with bigger balls trailing the path of the ball.
The wake in motion: two seconds of the first clip, frames 90 to 137, here at half the frame rate.

Drafts and tests

The first tile model had its seams turned by 90 degrees and too thin to survive at ball size. This is where it started, and the corrected version above.

Thirty black-and-white tiles in three lines of ten showing a ball with thin black seams and oval panels.
The first tile model: thin seams, and the ovals turned by 90 degrees. Six colourings, two to a line, five views of each.

Then the contrast: how many of the lightest balls to drop. Four versions of one frame, from the plain packing to contrast 10.

Animation cycling through four black-and-white basketball mosaics of the same frame, each emptier than the one before.
Four versions of one frame: the plain packing, then contrast 3, 6 and 10. More contrast leaves more of the white page empty.

Then the cut, and the wake, six frames a sixth of a second apart.

Six black-and-white basketball mosaic frames in two rows of three, with the top and bottom of each frame left empty.
The cut: the top 30% (upper row) or 20% (lower row) and the bottom 15% emptied, three frames each.
Six frames side by side of a black-and-white basketball mosaic, large balls trailing the moving ball.
The bold wake: six consecutive frames, a sixth of a second apart.

Sources

  1. Harmon, L. D. and Knowlton, K. C. (1969). Picture processing by computer. Science 164(3875): 19–29. doi:10.1126/science.164.3875.19
  2. Carion, N. et al. (2025). SAM 3: Segment Anything with Concepts. arXiv:2511.16719; ICLR 2026. arxiv.org/abs/2511.16719
  3. Zheng, P., Gao, D., Fan, D.-P., Liu, L., Laaksonen, J., Ouyang, W. and Sebe, N. (2024). Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research 3: 9150038. doi:10.26599/AIR.2024.9150038
  4. Floyd, R. W. and Steinberg, L. (1976). An adaptive algorithm for spatial grey scale. Proceedings of the Society for Information Display 17(2): 75–77.
  5. van der Walt, S. et al. (2014). scikit-image: image processing in Python. PeerJ 2: e453. doi:10.7717/peerj.453

Continue: Algorithmic Mosaic · How the algorithms developed, 1966–2026

Source image: Paolo Guidobono. Study, mosaics and animation: Charis Tsevis.

Charis Tsevis
Next

Studies in Perception I: A Document-Based Chronology, 1966–1969