Earlier this year, Google unveiled an eye-popping text-to-video games model called Genie that allowed to build basic playable games in 2D based on a short description. Six months on, AI in video gaming has made significant advances with gaming giants Roblox and Tencent presenting in recent days their own Generative AI technology that now allows to handle 3D environments and opens up the opportunity to reduce development time and costs by automating content creation.
At its annual developer conference, Roblox announced a generative AI model aimed at creating 3D environments and at modifying them in the blink of an eye (e.g. switching to a desert scene from a forest one, to a daytime scene from night, or adding a tree in a street…).
Roblox’s AI model is similar to large language models. It treats the 3D blocks that make up its virtual world as units that can be assigned a numerical value on the basis of how likely they are to come next in a sequence. That said, the insufficient 3D data to train the model has been a challenge, sometimes creating incoherent 3D environments (e.g. a computer in the middle of a lake). This has led Roblox to use a second AI model based on 2D data that checks if the content produced by the 3D model is coherent or not and proposes corrections.
Tencent’s GameGen-O, which was developed on data from close to 150 open-world games, goes a step further and lays the foundations for full AI game engines as it produces not only 3D dynamic environments but also characters simulation and complex actions.
Once again, the model has some limitations as the interactive control is restricted to simple movements, meaning that a player’s actions cannot influence the game world directly and that the video sequences produced by the model only simulate gameplay for now.
But clearly, Roblox and Tencent’s models represent a major step forward in the realm of open-world game generation based on AI and the current pace of innovation suggests that operational AI models should be available in a near future to both professional developers and non-professional users (Roblox highlighted that users with minimal 3D skills will be able to create environments).
In terms of implications, we view AI as a two-edged sword for the gaming/Metaverse industry. In the short term, Generative AI is expected to take assistive technology to a new level, materially reducing development times not only for gaming assets creation but also for music generation and animation, hence materially speeding up the creation of games and virtual worlds. It is also a game changer in terms of development costs as it is expected to make creative and development teams much more efficient by letting them focus on complex tasks (design of key avatars/settings, scenario, virtual world mechanics…) and get rid of basic artwork and time-consuming processes (simple code snippets, software testing and documentation…).
As an illustration, Electronic Arts commented recently that it expects to become 30% more efficient over the next 3 years thanks to AI, suggesting that margins should keep materially improving across the industry.
In the longer term, we see some downside risks as AI is expected to significantly lower the barriers to entry and democratize game development. This means that smaller teams (or even individuals) could create complex games with limited resources and take market share at the expense of established game makers.
These AI initiatives are also a concern for game design software firms including Unity and privately-held Epic as it’s likely that, at some point, some of the most powerful AI models will provide similar capabilities and will compete head to head with these firms’ own AI products.
