Game Level Blending with User Input
An AI that generates blended platformer levels from text prompts, mixing the styles of Super Mario Bros., Kid Icarus, and Mega Man into novel, playable levels.
We fine-tuned distilGPT-2 with a Mixture of Experts (MoE) architecture — one expert per game — so each expert learns that game’s level structure and style. During generation, activating multiple experts blends their knowledge, producing levels that inherit characteristics from each source game. Text prompts like “many solids, few hazards” are encoded via BART and injected through cross-attention, giving users direct control over what gets generated.
Level Blending
- 4 experts trained independently on Super Mario Bros., Super Mario Bros. 2, Kid Icarus, and Mega Man
- Each game contains levels of varying orientations and landscapes
- Activating multiple experts measurably shifts the expressive range of outputs toward a blend of the source games
- 98.6% of generated SMB segments were playable; 90% of full levels completed end-to-end
- Blended SMB + KI levels produced segments playable under both games’ movement rules (30.5%)
User Input & Control
- Type a prompt describing the level feel — obstacle density, openness, difficulty — and the model steers generation accordingly
- Prompt following improved +11% over random on hazard and moving platform tiles
- Adjust expert weights at inference time to prioritize one game’s style over another
Collaboration: Haoyu Chen, Ian Gauk & Zhikai Chen.