This experimental open-source AI turns prompts into playable Marvel, Star Wars and Harry Potter games

AI researchers think they're getting closer to being able to create video games from text prompts. A team at the Chinese University of Hong Kong's MMLab has released an open-source model that it claims is the first AI that can build a complete playable game from scratch (you might want one of the best laptops for game development).

According to the researcher paper, a developer can describe a game using natural language, and OpenGame will structure the project, write the code, debug errors and generate a final browser-based game. The researchers built demos for an Avengers-style platformer, a Harry Potter card battle game, KOF-style quiz fighters and a survival game inspired by Netflix's Squid Game.

OpenGame takes a different approach from many other AI game development models. Rather than generating individual code snippets or assets to put into an existing pipeline, it's an autonomous framework that can complete the whole process of generating a game itself.

Article continues below

It uses a specialised model called GameCoder-27B, which was trained on game development workflows with reinforcement learning on actual code execution. It can already generate over 150 different game types, the researchers say.

OpenGame's process comprises several coordinated stages. It begins by identifying the type of game requested, establishes a project framework and technical design, and then generates the necessary assets and automatically implements code and testing.

An evaluation system called OpenGame-Bench runs the generated game and judges it based on health, usability and whether it matches the original prompt. Any errors are then corrected until the framework deems the game to be playable.

(Image credit: CUHK MMLab)

The framework has an evolving ability set called Game Skill. Template Skill builds five types of template libraries from previous game generation experience, while Debug Skill maintains a protocol of error signatures and fixes that it's verified. The model stories a history of patterns, designs and fixes across projects, allowing OpenGame to learn over iterations, recycle solutions and improve its reliability over time.

(Image credit: CUHK MMLab)

While the process and output appear to represent a leap from previous models, many developers have doubts about how revolutionary OpenGame really is. One shot games shared online might look good, but there's no persistent state or multiplayer sync, and the resource cost of producing a game of any size could be huge.

"The demo is impressive, but I want to see it iterate with playtesting feedback instead of one-shot generation. Curious what breaks first on slightly longer game loops," one person writes on X.

"Demo games are the easy slice; canvas plus 20 lines of JS gets you Red Light Green Light in an afternoon," another person points out. "Real lift is state sync once two players join. Every agent coding project I've seen hits that wall around day three".

You can find OpenGame on Github

TOPICS
Joe Foley
Freelance journalist and editor

Joe is a regular freelance journalist and editor at Creative Bloq. He writes news, features and buying guides and keeps track of the best equipment and software for creatives, from video editing programs to monitors and accessories. A veteran news writer and photographer, he now works as a project manager at the London and Buenos Aires-based design, production and branding agency Hermana Creatives. There he manages a team of designers, photographers and video editors who specialise in producing visual content and design assets for the hospitality sector. He also dances Argentine tango.

You must confirm your public display name before commenting

Please logout and then login again, you will then be prompted to enter your display name.