We recently created one of the most exciting and complex campaigns for our client, the E.ON Hungária Group, which has been a dedicated supporter of Hungarian water polo since 2014. Water polo is a national treasure, and like the millions of fans and E.ON, we are passionate about it. So, when it became clear that artificial intelligence, which will shape our future, had no idea about this beautiful sport, we decided to smarten it up and teach it how to play water polo.

Currently, AI doesn't understand the unique dynamics of the sport (moving in water, ball handling, specific actions), so it can't generate accurate images of it. Even the most advanced image-generating models fail when it comes to water polo, imagining the sport with players in ankle-deep water, multiple balls, floating goals, and players wearing helmets jumping out of the pool. So, how can we influence the development of AI from here in Hungary? With dedicated, AI-savvy colleagues, good organization, and a lot of hard work and perseverance!

‍

Implementation

ACG recognized the potential of artificial intelligence early on, so a year ago, we established a dedicated AI division and placed a strong emphasis on training our colleagues. Thanks to this, we quickly found the right partner. To fix the AI's blind spot, we enlisted the help of Tengr.ai, a team with Hungarian ties who specialize in AI image generation and are among the best in the world. The ACG team first started the AI's training with a theoretical crash course, summarizing the basics of the sport in detail for the AI.

Next, we began compiling a massive training dataset. To teach the AI to recognize and realistically portray the specifics of water polo, we couldn't just create a simple collection of "water polo pictures." We had to build a structured curriculum that teaches the AI individual concepts separately, then allows it to creatively combine this knowledge.

For this, the ACG team sifted through the archives of our partner, the Hungarian Water Polo Federation, and selected thousands of images for the campaign. With the help of our colleagues who are active water polo players, we clustered and tagged these images so the AI could easily learn the sport from them. To get an even more perfect result, in addition to the archival materials, we organized a special shoot at the Alfréd Hajós Swimming Stadium, where we created film and photo tutorials with the help of professional players, recording the typical movements of offensive, defensive, and goalie positions both above and below the water.

This was followed by a lengthy training process, during which we collaborated with the Tengr.ai development team to refine the accurate depiction of players' clothing and accessories (e.g., cap type and number, chest number), postures and movements (e.g., spin shot, ball handling, blocking), and water polo equipment (e.g., goal shape, field boundaries).

‍

The Result

After nearly two months, we have now reached a point where we can share the results with users. Today, the open test of the water polo image generation model has launched on the tengr.ai website, where anyone can now see how much more the AI knows about water polo than it did a few weeks ago. We are proud that, in addition to the two silver medals, by smartening up the AI, we have also contributed to raising the prestige of Hungarian water polo. We can now say, along with our partners, E.ON and Tengr.ai, that we Hungarians have not only taught the world how to play water polo, but we have also taught artificial intelligence!

‍