Fei-Fei Li’s World Labs speeds up the world model race with Marble, its first commercial product

World Labs, established by AI visionary Fei-Fei Li, has unveiled its inaugural commercial product, Marble, which is now accessible through both freemium and paid subscription models. This innovative tool allows users to convert text prompts, images, videos, and even 3D layouts into editable and downloadable 3D environments. Marble's launch follows a limited beta preview that started just two months ago, marking a significant advancement for World Labs since its debut last year, where it secured $230 million in funding. This places the startup in a competitive position against other entities developing similar world modeling technologies. World models are sophisticated AI systems that create an internal representation of environments, enabling predictions of future scenarios and action planning. Unlike competitors such as Decart and Odyssey, which offer free demos, or Google’s Genie, currently in its research phase, Marble stands out by generating consistent, downloadable 3D environments instead of transient worlds that morph as users explore. This unique feature minimizes inconsistencies and allows users to export their creations in formats like Gaussian splats, meshes, or videos. Marble is groundbreaking in that it introduces AI-native editing tools alongside a hybrid 3D editor that lets users define spatial structures before the AI enhances the visual details. Justin Johnson, a co-founder of World Labs, emphasizes that this new model category is still evolving and is expected to improve continuously. The beta version showcased Marble’s ability to create interactive 3D scenes from single images, albeit with some limitations in exploration and occasional rendering errors. In testing the beta, I found that Marble could produce impressive environments based solely on image prompts, ranging from game-like settings to photorealistic depictions of familiar spaces. Although some edge morphing issues were noted, improvements have been made in the final launch. Johnson highlights that Marble aims to empower users with creative control. The full launch allows for the uploading of multiple images or short videos, enabling the generation of realistic digital twins of spaces. An experimental editor named Chisel facilitates the creation of basic spatial layouts, allowing users to guide the visual style with text prompts, thus separating structure from aesthetics. New features further enhance user control, including the ability to expand generated worlds. Users can instruct the model to generate additional details in specific areas as they explore. Additionally, for those looking to create expansive environments, there is a “composer mode” that allows for the merging of multiple worlds. Marble is offered through various subscription plans, ranging from free access, which permits a limited number of generations, to a comprehensive $95 Max tier that provides full features and extensive generation capabilities. Johnson anticipates that the initial applications for Marble will be found in gaming, visual effects, and virtual reality sectors. Despite some skepticism within the gaming community regarding the impact of generative AI on the industry, Johnson believes Marble can be a valuable tool for game developers, assisting in the creation of background environments that can be imported into game engines like Unity or Unreal Engine. Furthermore, for visual effects work, Marble promises to address the inconsistencies often seen in AI-generated videos, providing artists with precise control over scene staging and camera movements. While the focus is not currently on virtual reality applications, Marble is compatible with popular VR headsets like Vision Pro and Quest 3, making it possible to experience generated worlds in a virtual setting. The tool also holds potential for use in robotics, as it aids in simulating training environments, thereby advancing the capabilities of robotic systems. Fei-Fei Li envisions Marble as a pivotal step toward developing a spatially intelligent world model. She believes that the next evolution of world models could enable machines to achieve unprecedented levels of spatial intelligence, bridging the gap between AI's ability to read and write and its capacity to understand and manipulate three-dimensional spaces.