Marble

Genie 3 is DeepMind’s next-generation, general-purpose world model capable of generating richly interactive 3D environments in real time at 24 frames per second and 720p resolution that remain consistent for several minutes. Prompted by text input, the system constructs dynamic virtual worlds where users (or embodied agents) can navigate and interact with natural phenomena from multiple perspectives, like first-person or isometric. A standout feature is its emergent long-horizon visual memory: Genie 3 maintains environmental consistency over extended durations, preserving off-screen elements and spatial coherence across revisits. It also supports “promptable world events,” enabling users to modify scenes, such as changing weather or introducing new objects, on the fly. Designed to support embodied agent research, Genie 3 seamlessly integrates with agents like SIMA, facilitating goal-based navigation and complex task accomplishment.

Atlas is a next-generation omni world model for spatial intelligence that natively operates on text, images, video, and 3D. Built as a multimodal autoregressive diffusion transformer, it combines inputs into a shared spatial context and generates what comes next while staying consistent in 3D with everything it has seen and imagining what lies beyond it. Atlas supports world generation, reconstruction, and simulation across a broad range of tasks. It can generate images and videos from one or more reference images with pixel-perfect camera control, producing long, coherent videos with manually designed camera paths. For spatial reconstruction, Atlas can recreate real-world scenes from sparse input images, generate novel views, and produce explicit 3D outputs such as point clouds and 3D Gaussian splats. More input views provide additional context, allowing the model to reduce imagination and create increasingly faithful reconstructions. Atlas also models how worlds evolve over time.