World's First AI Native UGC Game Engine Launched! Instantly Create GTA-Style Worlds with Text Input, Experience the Demo Now}

Editor | ScienceAI

"From now on, the future of gaming will no longer be solely crafted by professional designers, but will be created and experienced by everyone in real-time."

Today, the world's first real-time world model-driven AI native game engine has been unveiled!

The system is called Mirage, developed by Dynamics Lab.

This engine is designed for building dynamic, interactive, and continuously evolving gaming experiences. Players can generate and modify the entire game world through natural language, keyboard, or controller inputs in real-time.

Mirage supports multiple game types.

Currently, two playable demos are available: City Brawl (GTA style) and Coastal Drift (Extreme Racing Horizon style).

• City Brawl: https://demo.dynamicslab.ai/chaos
• Coastal Drift: https://demo.dynamicslab.ai/drift

All scenes are generated in real-time, not pre-scripted. We see a dynamically evolving interactive world that changes with player actions.

      ^{City Brawl (GTA style)}

      ^{Coastal Drift (Extreme Racing Horizon style)}

The team at Machine Heart tested City Brawl, which features a control interface with options on the left and street view on the right, as shown below:

After a brief experience, they found the game still has high latency, and the responsiveness of character movements can be further improved. Zhiting Hu shared on 𝕏 (Twitter) that network latency is likely the main cause:

Mirage also supports users uploading initial images to start gameplay. The team tested with a nearby photo, and the video is shown below:

Although the current experience still has gaps compared to expectations, Mirage represents a frontier breakthrough in AI technology.

Compared to recent achievements like Google’s AI Doom/Genie, Decart’s AI Minecraft, and Microsoft’s AI version of "Thor II," Mirage has three major advantages:

• Real-time UGC creation: Supports content generation during gameplay via text input, a significant breakthrough in world model-based game generation.
• Film-quality visuals: Achieves photo-realistic visual effects beyond early pixelated systems.
• Persistent interactive experience: Supports over ten minutes of continuous interaction, greatly expanding the boundary of generative gameplay.

Netizens in the comments have expressed high praise and anticipation for Mirage.

Those eager to try it out should act quickly, as the servers might already be near collapse.

The development team firmly believes that with continuous breakthroughs in real-time generation technology, Mirage will reshape the future landscape of the gaming industry.

UGC 2.0: The Rise of Generative Gaming

Traditional games rely on pre-designed maps, fixed city layouts, scripted missions, and finite experiences. Mirage breaks these boundaries, allowing players to create new experiences in real-time during gameplay.

Using natural language commands, keyboard inputs, or controllers, players can generate escape routes, summon vehicles, or expand city skylines. The system responds instantly, seamlessly integrating user-generated elements into a continuously evolving simulated world—making the game not just interactive but co-evolving with players.

Mirage supports various game genres, from racing and RPG to platform jumping:

This is the new world of UGC 2.0:

• Zero-barrier creation: Anyone can generate their own game world with just text descriptions.
• Real-time co-creation and evolution: Players create, evolve, and reshape game content instantly during gameplay.
• Infinite dynamic experiences: Each session is unique, never repeating, fully abandoning pre-scripted content.

Built on Next-Generation AI

Mirage is based on cutting-edge World Model research, utilizing a Transformer-based large-scale autoregressive diffusion model capable of generating controllable high-fidelity video game sequences. The framework integrates multiple innovations:

Game-based foundational training

It is built on a robust training foundation—aiming to understand and generate rich gaming experiences. This begins with large-scale collection of diverse game data from the internet, capturing various mechanics and styles.

To this end, Dynamics Lab developed a dedicated data recording tool to capture high-quality human gameplay recordings. These meticulously curated samples enrich the dataset, enabling the model to learn complex player behaviors and game logic.

Subsequently, the data is fed into a specialized training pipeline focused on game-specific learning, allowing the model to internalize game rules, system patterns, and interaction dynamics.

Ultimately, the model can generate coherent, realistic, and flexible game content, breaking traditional pre-scripted limitations.

Interactive generation and real-time control

By integrating frame-level prompts into its core, Mirage redefines real-time interaction. Players can input commands or prompts that are immediately interpreted during gameplay.

Supports cloud gaming for anytime, anywhere play:

• Dynamic input system: Handles player inputs with ultra-low latency via keyboard and text, enabling near-instant responses.
• Real-time output: Visual updates are transmitted through a full-duplex communication pipeline, ensuring smooth interaction without delay.

Technically, Mirage is based on a customized causal Transformer model, enhanced with:

• Dedicated visual encoder
• Improved positional encoding
• Optimized structure for extended interaction sequences

The system combines the strengths of LLMs and diffusion models, supporting high-quality, coherent content generation. Distillation strategies are used to balance speed and quality.

Players can reshape environments with natural language commands, triggering instant world updates. It features a KV cache-driven long context window to maintain visual consistency amid real-time world evolution.

Key features

• 16 FPS at SD resolution for real-time interaction
• Dynamic UGC: Generate content via natural language commands
• Extended gameplay: Minutes-long interactive sessions with visual consistency
• Cloud streaming: Cross-platform instant gaming without downloads
• Infinite replayability: Unique experiences each time
• Multi-mode control: Text, keyboard, and controller inputs supported

Team and future outlook

The Mirage team comprises experienced AI researchers, engineers, and designers from Google, NVIDIA, Amazon, Sega, Apple, Microsoft, Carnegie Mellon University, and UC San Diego.

They share a vision: "Generative gaming is not just a feature but a new medium. Mirage leads a future where games are imagined, inspired, and experienced, no longer requiring downloads or pre-designed content."

^{Blog: https://blog.dynamicslab.ai/}