90% Faster, 100% Code-Free: MLLM-Driven Zero-Code 3D Game Development

UniGen is the first end-to-end multi-agent framework for automated Unity game creation. From a natural language description, UniGen generates runnable 3D games without requiring a single line of manual code.

Backed by Multimodal Large Language Models (MLLMs), UniGen automates the entire pipeline:

1. Planning Agent – interprets natural language into structured blueprints.
2. Generation Agent – produces robust Unity C# scripts.
3. Automation Agent – builds and configures Unity scenes.
4. Debugging Agent – fixes issues interactively via conversational debugging.

🎥 Demo Video
📄 Paper

• Zero-Code Development – create Unity games with just natural language prompts.
• End-to-End Automation – no manual engine configuration, component binding, or script integration.
• Multi-Agent Framework – specialized agents for planning, code generation, scene assembly, and debugging.
• High Efficiency – reduces development time by 91.4% compared to manual workflows.
• High Functional Completeness – achieves >89.5% correctness across diverse game prototypes.
• Unity Integration – outputs a ready-to-run Unity project with Editor + Runtime scripts.

• Python 3.8+
• Unity 2021+ (tested with LTS versions)
• An OpenAI / MLLM API key (e.g., GPT-4.1 or Gemini-2.5)

1. Describe your game idea in plain English
2. Run the pipeline:bash python blue_pipeline.py
3. UniGen will: a)Generate Unity C# scripts (PlayerController, UIManager, etc.). b)Build a scene with configured objects and components. c)Save all assets into UnityProject/Assets/. d)Open the Unity project, press play, and your game is ready!
4. open the Unity project, press play, and your game is ready! 🎮

UniGen/
│── blue_pipeline.py         # Main automation pipeline
│── gpt_interface.py         # GPT/MLLM connector
│── requirements.txt
│── README.md

UniGen was tested on three prototypes:

• Average development reduced from 140 mins → <12 mins.
• Manual operations reduced from 75 → <5.

UniGen’s pipeline operates via four agents:

• Planning Agent → JSON blueprints of entities & logic.
• Generation Agent → Unity-ready C# scripts.
• Automation Agent → Scene construction via Unity Editor APIs.
• Debugging Agent → Conversational, automated error correction.

• Complex NPC behavior trees not yet fully automated.
• Multiplayer synchronization requires future enhancements.
• Currently optimized for Unity, but engine-agnostic by design.