Google DeepMind releases Genie 4, a world model users can change from inside the scene

AI & Compute

Google rolls out Genie 4 with editable worlds and native audio

Google DeepMind’s Nov. 18 rollout gives Ultra subscribers editable, audio-enabled worlds, while Genie 3 gains 3-minute generations, panoramas, linked journeys and Gemini assistance.

By
Rebecca Stone
Filed
Channel
AI & Compute
Read
3 min read

Google DeepMind began rolling out Genie 4 on Wednesday, Nov. 18, to eligible Google AI Ultra subscribers through Project Genie, its experimental prototype for generating and exploring interactive worlds. Google AI Pro members can join a waitlist, so the first release is tiered rather than universally available.

What can users change in Genie 4?

Users do more than set conditions before a world starts. Google said: “Users can type a new prompt while navigating a world to shift the weather, drop in a character or redirect the action.” The prompt operates inside the running scene, changing its setting, cast or direction without requiring the user to restart from the original instruction.

Audio follows the same model-driven approach. Google described Genie 4 as “the first version to pair its visuals with real-time native audio that adapts to the scene,” ranging from ambient sound to character dialogue. The announcement offered no latency measurement for that audio.

Together, those features turn scene generation into an ongoing interaction. Weather, characters and action can change while the user moves through the world, while generated sound adapts to what is happening on screen. The system accepts new instructions instead of stopping at one predetermined sequence.

What changed in Genie 3?

Genie 3 remains available in the model picker, but Google also gave it upgrades aimed at longer sessions and broader starting points. One generation can now run for up to 3 minutes, up from 60 seconds, and users can link several worlds into one journey.

Genie 3 can also begin with a 360-degree panorama image and accept natural-language prompts during a session. The 3-minute generation window is the disclosed Genie 3 ceiling; Google did not state Genie 4’s maximum session length.

Both models therefore remain in the picker while Google tests different capabilities. Genie 3’s changes center on sustained journeys and panoramic starts; Genie 4 introduces live scene mutation and adaptive native audio.

What can the agent and Gemini do?

A new Project Genie agent builds a world from a conversation. Google’s multimodal models generate the initial visuals and a custom background score from that exchange, providing a conversational route into world creation.

Gemini models can then join a session, watch what appears on screen, suggest where the user should go next and trigger an in-world event. Those functions place Gemini inside the navigation loop rather than limiting it to an external chat window.

A community gallery completes the sharing layer. Users can publish worlds they create and remix worlds made by others, allowing one generated environment to become the basis for another.

What remains unproven?

Google did not disclose Genie 4’s resolution, frame rate or maximum session length. Those omissions matter for judging an interactive model whose pitch rests on real-time response, because they leave the raw rendering envelope and session ceiling unstated.

The company also said Genie 3 and Genie 4 remain frontier research. Its remaining technical work spans camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence—four areas that determine whether a generated world stays coherent as it changes.

Watermarking and screening already surround the prototype. All video and audio generated in Project Genie carries a SynthID watermark, while model inputs and outputs pass through safety screening. Those controls apply across the release, including the agent, Gemini assistance and published community worlds.

The commercial picture remains defined by access. Genie 4 starts with eligible Ultra subscribers, Pro members wait, and Google’s announcement provides no wider release schedule. The next commercial marker is whether access expands beyond those tiers as Google works to close the camera-latency, physical-interaction, audio-precision and scene-permanence gaps it acknowledges.

Source: mlq.ai

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,131 articles

Related articles

« Previous articleNext article »