Genie 4 Launches with Interactive World Simulation and Video Controls
Google DeepMind launched Genie 4, introducing an interactive world model that generates controllable environments from text prompts and single images. Virtual production teams and creators can now test camera trajectories and physics before committing to final video renders.
Google DeepMind launched Genie 4, introducing real-time interactive world modeling into the ecosystem that powers video generation tools like Google Veo 3. The model enables creators to generate steerable, action-driven virtual environments directly from text prompts, sketches, or still images. This release shifts generative video production away from passive clip generation toward playable virtual spaces that filmmakers can direct interactively.
What's new
Genie 4 functions as a foundation world model, generating continuous visual environments that respond to user action inputs in real time. While standard diffusion video models calculate fixed timelines, Genie 4 maintains spatial consistency and scene physics as an agent or virtual camera moves through the space.
Key capabilities introduced in Genie 4 include:
- Action-controllable navigation: Users can steer camera angles and movement paths across generated environments using directional inputs.
- Single-image environment generation: Creators can input a single concept frame—from Google Veo 3, Midjourney v6, or photographic plates—and expand it into a traversable scene.
- Memory and spatial consistency: The model tracks scene geometry across extended navigations, reducing the visual drift and warping common in standard generative video.
- Physics and collision modeling: Genie 4 calculates gravity, object boundaries, and spatial perspective on the fly without relying on an external 3D game engine.
How it fits your workflow
For directors, VFX supervisors, and animators, Genie 4 changes the previs and blocking stages of production. Traditional pre-visualization requires building rough low-poly environments inside Unreal Engine, Blender, or Maya before testing camera lenses and actor blocking. With Genie 4, a production team can generate a navigable location from a written treatment, steer a virtual camera through the set to scout angles, and export the resulting passes.
This workflow addresses the primary limitation of standard text-to-video generators. While platforms like Runway Gen-3 Alpha, Kling 2.0, and OpenAI Sora produce photorealistic clips, their camera controls rely on prompt phrasing or rigid trajectory presets. Genie 4 provides an interactive sandbox where camera operators test dynamic coverage continuously before committing compute resources to high-resolution final renders in Google Veo 3.
The model also gives game designers a fast prototyping pipeline. Concept artists can feed 2D character sheets and background art into Genie 4 to test platforming physics, lighting responses, and environmental scale within seconds.
What it costs / how to try it
Google DeepMind has released Genie 4 as a research preview, with access currently limited to select academic partners and enterprise creative teams via the Google DeepMind research portal. Broader commercial availability and direct integration into Google Cloud Vertex AI and Google Veo 3 pipelines have not yet received formal rollout dates.
Read the original announcement on Google Veo 3 ↗