Neural network clusters Steam games by single screenshots

Polygon reports that in late December 2025 a developer using the handle Newbie Indie Game Dev downloaded the first Steam screenshot from thousands of games and fed them into a neural network to test whether a single screenshot could predict a game's success. The developer pulled whichever image appeared first in each game's Steam gallery, pared the set to 10,000 titles (including every game with at least 3,000 reviews plus a randomized selection), and added metadata such as tags and prices.

All of that was passed to EfficientNet-B3 to produce an interactive clustering called The Gaming Map. The analysis found that the most-reviewed games tended to be action-packed 3D titles — screenshots emphasizing guns, people in motion, war and explosions — and that many of those images leaned toward warmer tones; many, though not all, were first-person shooters, military or tactical games.

Newbie cautioned against sweeping conclusions, noting exceptions and visually distinctive games that clustered far from the popular groups. He also noted that action-heavy 3D games are often AAA productions with bigger budgets and marketing, and that a single screenshot was a reasonable predictor of expected price — "Better graphics, higher cost." Newbie stressed that correlation does not mean causation and called the results preliminary: "I didn’t have high expectations going in...

Key Topics

Tech, Steam, Game Screenshots, Action Games, Aaa Publishers