Marble: A Multimodal World Model (www.worldlabs.ai)

AI startup Marble has announced general availability of Marble, a generative multimodal world model that builds editable 3D worlds from text, single or multiple images, short videos, or coarse 3D layouts. Beyond simple text-to-world generation, Marble supports multi-image and video stitching for view-consistent scenes, interactive 2D/3D editing, one-step regional expansion, and a composer mode to tile or join worlds into much larger spaces. The new Chisel tool lets advanced users lay out coarse geometry (boxes, planes, imported assets) to define structure and then apply text-driven styling—decoupling structure from appearance for finer control. Technically, Marble exports high-fidelity Gaussian splats (semisuperimposed particles) and triangle meshes including collider meshes for physics and high-quality meshes for visual fidelity, and it can render pixel-accurate videos or enhanced videos that clean artifacts and add dynamic elements. Marble integrates with Spark, an open-source THREE.js renderer, and supports iterative workflows—generate, edit, expand, compose, and export—making it easy to lift 2D image outputs from other generative models into cohesive 3D assets. The release, plus Marble Labs (a creative hub with case studies and tools), signals a practical step toward spatial intelligence, enabling new pipelines for gaming, VFX, robotics simulation, design, and interactive agent environments.