Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training
The paper’s core move is stripping HUDs out of gameplay videos so they can train cleaner world models.
Game2World pairs a UI taxonomy with G2WEngine, which extracts reusable interface assets from real gameplay and synthesizes coherent overlays on clean footage. The resulting dataset includes 96K synthetic paired videos and 1,079 in-the-wild clips from 303 games. The authors also introduce GameCleaner, a mask-free UI removal model that preserves scene content while removing HUD elements. In a pilot, world models trained on UI-free gameplay scored 6.83% higher on VideoReward than models trained on UI-overlaid data. HF Daily Papers' note
Game2World pairs a UI taxonomy with G2WEngine, which extracts reusable interface assets from real gameplay and synthesizes coherent overlays on clean footage. The resulting dataset includes 96K synthetic paired videos and 1,079 in-the-wild clips from 303 games. The authors also introduce GameCleaner, a mask-free UI removal model that preserves scene content while removing HUD elements. In a pilot, world models trained on UI-free gameplay scored 6.83% higher on VideoReward than models trained on UI-overlaid data. HF Daily Papers' note
score 5