Genie 4 is Google DeepMind's natively multimodal world model, announced on October 27, 2026, and genie4.art is an independent guide that explains it in plain language.
What the guide covers:
- What Genie 4 actually is, and why "world model" is a different category from text-to-video
- A side-by-side comparison with Genie 3, using the same dressing-room scene from both models
- The three-layer architecture: Gemini Omni grounding visuals, Gemini 3.8 Flash running a real-time observe-plan-replan loop, and a renderer that recomputes the world continuously
- Real-time steering, physics and character realism, native audio, promptable world events, and SynthID watermarking
- Practical use cases: film and TV pre-visualisation, game prototyping before engine work, robotics and agent training, and bringing archival images to life
- Access and availability: trusted testers, AI Ultra subscribers via Project Genie, plus Gemini API and AI Studio partners
- A prompt composer that helps you set a place, describe motion, name the audio, and leave room to steer afterwards, with ready-made example prompts
Who it is for: filmmakers and studios exploring pre-visualisation, game designers validating mechanics early, robotics and agent-training teams, and anyone who wants to understand the shift from generated clips to playable worlds.
The site is free, requires no account, and is updated as DeepMind publishes new demos and details. Read the full guide at https://genie4.art
