Multiverse: Language-Conditioned Multi-Game Level Generator

🌟 Overview

Multiverse is a language-conditioned multi-game level generator that enables both multi-game level generation and cross-game level blending through textual specifications.

Key highlights:

• 🎮 Multi-game generation — a single model covers multiple game domains
• 🔀 Cross-game level blending — smoothly interpolate between games via latent space
• 📝 Language-guided control — use text prompts to specify structural characteristics
• 🔗 Shared latent space — aligns textual instructions and level structures across games
• ✨ Zero-shot compositional generation — generate novel levels from compositional text prompts

Unlike prior text-to-level generators limited to a single game domain, Multiverse learns a shared representation that captures structural relationships across game domains, enabling controllable blending through latent interpolation.

Compositional Text Blending	Embedding Interpolation Blending

---

📦 Dataset

5,576 annotated levels from human-authored game levels, collected from VGLC and PCGRL environments.

	🍄 Super Mario Bros	🗡️ The Legend of Zelda	🏃 Lode Runner	🏰 Dungeon
Level
Instruction	flat ground with pipe right and question blocks above center	spiral wall corridors around central floor with water pockets	ground platforms top and mid ladders left and center rope top gold right	scattered walls forming winding paths with bats in open areas
Annotations	1,240	1,836	1,200	1,300

The dataset includes the original instructions (annotation.csv), a generalized version (generalized_annotation.csv), and a blended version combining instructions from two different games (blended_instruction.csv).

To set up the dataset:

bash setup_dataset.sh

---

🛠️ Environment Setup

🐳 Build Docker Image

docker build -t multigame .

---

🚀 How to Run

🐳 Using Docker

The run_docker.sh script automatically finds and allocates an available GPU.
Container names follow the format: multigame_gpu{GPU_NUMBER}_{DATETIME}.

Basic Usage

./run_docker.sh <command> [args...]

🔑 Weights & Biases Setup

Create a .env file in the root directory:

WANDB_API_KEY=your_wandb_api_key

---

🏋️ Training

🌐 Multiverse (Full Model)

Trains with the game-general loss (gen). Runs over seeds 0–9.

./run_docker.sh python train.py gen_threshold=0.3 loss_weights.base=0.0 loss_weights.gen=1.0 tsne_interval=200 render_interval=200 vit_eval_freq=200

🔩 Single-positive Contrastive Learning (Baseline)

Trains with the base loss only. Runs over seeds 0–9.

./run_docker.sh python train.py overwrite=true loss_weights.base=1.0 loss_weights.base=0.0 tsne_interval=200 render_interval=200 vit_eval_freq=200

📝 Text-Blend Evaluation

Evaluates cross-game level blending guided by compositional text prompts using a trained checkpoint.

./run_docker.sh python eval_textblend.py seed=0 \
  checkpoint_path="saves/e2e_exp-def_clipdr-0.1_cliptemp-0.14_gen-1.0_s-0/epoch_2000/checkpoints/checkpoint_epoch_2000.pt" \
  checkpoint_epoch=2000

---

📄 License

This project is licensed under the MIT License.