Arcimboldo's Style
The task was to approximate a target image using a collection of cliparts. Participants had to choose which cliparts to place, in what order, and at which coordinates, so that the final canvas would be as close to the target as possible.
Let's take a closer look at the tests.
The First Tests
The first test was exact: every team got 200 points for it.
The second test was also not very hard to solve optimally, and every team found the best answer.
Random Letters
The third test stayed competitive until the very end. Several teams kept improving the best answer and repeatedly pushed the others far behind.
The test was constructed like this: 2000 random letters were put onto the canvas. At least one pixel was visible for 1097 of them, but participants were allowed to place only 1000 cliparts. The intended approach was to reconstruct about 1100 letters with zero error, then choose the most important ones. Since the total error became very small, every tiny improvement had a significant effect on score distribution.
Colored Dots
The next group of tests used colored dots. The dots were small, and there were many of them, so this was almost pixel-by-pixel approximation, just with thick pixels.
Look closely at Katarina's eyes and at the small stripes on her sleeve. They are drawn with the edges of cliparts. The same technique also appears in other tests with many cliparts, such as the Wave.
For the Wave, you can also watch a short video showing how the best team's solution places the cliparts on the canvas one by one, mostly from bottom to top.
Fewer Cliparts
In another group of tests, only a small number of cliparts could be placed, but the cliparts themselves became much easier to see. The result no longer looks like dense rasterization: individual leaves and flowers become part of the composition.
The Guest of Honor
Finally, there was a test with the guest of honor himself: Arcimboldo. Just enjoy the portrait. For this page, we also recolored the participants' solution using the colored vegetable cliparts.
You can browse the best output of every team on every test in the results gallery.
So, How Do You Solve It?
Honestly, we do not know. This was an output-only task, and in many teams the solutions were produced by agents, so the participants may not know either. It looks like agents are seriously breaking the old assumptions behind marathon and optimization contests. Maybe the genre is not dead yet, but it is clearly in trouble.
Still, looking at the final outputs, the following ideas seem useful:
- Greedy search, local search, beam search, and simulated annealing are natural baselines: keep the current canvas, try adding, moving, or replacing a clipart, quickly estimate how the error changes, and repeat.
- In tests with many small cliparts, treat the task almost like rasterizing the image with thick pixels.
- Drawing with clipart edges is powerful: many fine details, such as eyes, sleeve stripes, and contour lines, are made from barely visible edges of larger cliparts.
- It helps to preselect useful cliparts from the large collection and search mostly among them. For example, the best Mona Lisa with leaves output uses only 35 distinct cliparts, not the whole set of 326 leaves.
- It may help to detect high-contrast and low-contrast regions of the target image, and spend more cliparts on high-contrast regions. This one is less certain.
- It helps to improve already good solutions instead of restarting from scratch every time.
- And finally: do not forget to submit the best solutions you have found before the contest ends.