Genie 4 Introduces Real-Time World Simulation and Interactive Control
Google DeepMind unveiled Genie 4, adding real-time interactive simulation and consistent physical dynamics to its video model architecture. The shift gives game designers, animators, and VFX artists controllable 3D environments rather than unguided video clips.
Google Veo 3, Google DeepMind's generative video platform, gained a foundational expansion with the introduction of Genie 4, a real-time world simulation model designed to handle interactive environments and continuous physical dynamics. Unlike standard text-to-video generators that produce fixed, non-interactive clips, Genie 4 computes physics, collision, and camera trajectories continuously based on user input. This development bridges the gap between pre-rendered video in Google Veo 3 and interactive, spatially coherent 3D scenes.
What's new
DeepMind structured Genie 4 to advance past the limitations of purely visual diffusion models by combining video prediction with action-conditioned world simulation. Key updates include:
- Real-time interactive generation: Genie 4 generates coherent frames at interactive speeds, reacting to frame-by-frame user inputs without the rendering delays common in batch video workflows.
- Environmental persistence: The model maintains object permanency across extended generation sequences, preventing the morphing and sudden visual disappearances that occur when a camera pans away and returns.
- Unified visual pipeline: By connecting world-state simulation directly to the visual rendering stack developed for Google Veo 3, the system outputs accurate lighting, materials, and textures while adhering to strict spatial boundaries.
- Action conditioning: Users control camera vectors, object collisions, and agent navigation through direct text instructions, game-controller inputs, or defined motion paths.
How it fits your workflow
For virtual production supervisors, animators, and game artists, Genie 4 offers a method to prototype interactive sequences without assembling complex 3D assets in game engines first. Standard AI video generation tools like Runway Gen-4 and Kling 2.1 deliver finished video files with limited real-time steering. In contrast, Genie 4 operates as a dynamic sandbox where creators can explore environments, test shot compositions, and refine camera blocking continuously.
Directors can use Genie 4 to scout virtual sets and evaluate complex camera moves before sending scenes for final rendering in Google Veo 3 or Sora 2. Instead of generating dozens of separate prompt variations in tools like Pika 2.2 or Luma Dream Machine to find a workable angle, an operator navigates through the scene directly. The generated simulation can serve as an animatic, a previs sequence, or a spatial reference pass for downstream VFX work.
This simulation architecture also positions Google against emerging spatial AI platforms like World Labs. By anchoring video generation to simulated physical rules, Google enables creators to generate complex physical events—such as falling debris, fluid dynamics, and terrain navigation—with predictable behavior that standard generative video systems routinely fail to maintain.
What it costs / how to try it
Google DeepMind is currently previewing Genie 4 through technical research access, with planned deployments across Google Cloud Vertex AI to support generative media workflows alongside Google Veo 3. Google has not yet announced commercial pricing tiers or general API availability dates.
Read the original announcement on Google Veo 3 ↗