Google rolls out Genie 4 for interactive world generation
Google has released Genie 4, its most realistic interactive world model yet, for Google AI Ultra subscribers.
2 min read, Thursday, 19 November

Google is rolling out Project Genie updates for eligible subscribers, featuring the new Genie 4 model, a creative agent, and a community gallery. The platform combines advanced world models with Gemini models to let users generate, explore, and shape interactive digital worlds directly from a web browser.
Key facts
- Genie 3 session length
- 3 minutes
- Visual fidelity
- Genie 4: Highest fidelity
- Genie 3: —
- Max session duration
- Genie 4: —
- Genie 3: 3 minutes
- Panorama grounding
- Genie 4: —
- Genie 3: Supported
| Metric | Genie 4 | Genie 3 |
|---|---|---|
| Visual fidelity | Highest fidelity | — |
| Max session duration | — | 3 minutes |
| Panorama grounding | — | Supported |
Genie 4 capabilities
Genie 4 delivers high-fidelity visuals, physics, and native audio. Users can type natural language prompts mid-session to change environments in real time. Shifting the weather, adding characters, or redirecting action happens instantly while navigating. The model also introduces real-time native audio that adapts as the scene unfolds, providing ambient soundscapes and character dialogue.
Genie 3 updates
While Genie 4 serves as the most realistic offering, Genie 3 remains available for longer runs and panorama-seeded spaces. Recent updates to Genie 3 include:
- Extended sessions: Explore for up to 3 minutes per generation, an increase from 60 seconds.
- World chaining: Link multiple worlds together for longer journeys.
- 360-degree panorama grounding: Seed new worlds from panorama images.
- Real-time promptable events: Dynamically change the environment with text while exploring.
Creation and community
The new Project Genie agent helps users build custom universes through natural language conversation. The agent coordinates Google's multimodal models to establish initial visuals and custom background soundtracks. Additionally, Gemini models can join sessions to observe screens, suggest next steps, or trigger in-world events automatically.

Users can publish their creations to a new community gallery to share and remix work. Safety measures include SynthID watermarking for all generated video and audio, alongside safety screenings for inputs and outputs.
Source: Original announcement
Worth a read?



Comments