Skip to content
Bud's Wiki

Project Genie: Experimenting with infinite, interactive worlds

TL;DR

Section titled “TL;DR”

Google DeepMind has rolled out Project Genie — a Google Labs web prototype powered by the Genie 3 world model — to Google AI Ultra subscribers in the US. The system lets a user describe a world via text and images (assisted by Nano Banana Pro for “World Sketching”), step into it, navigate it in real time at 20–24 fps / 720p, and remix existing worlds. It is positioned as the productization step after the August 2025 Genie 3 research preview, and as a step toward AGI agents that can reason in the diversity of the real world. The post is product framing rather than a technical report: no architecture details, no benchmarks, no comparisons to prior systems.

Key claims

Section titled “Key claims”
  • Project Genie is rolling out to Google AI Ultra subscribers (US, 18+) starting Jan 29, 2026, as an experimental Google Labs prototype [§rollout].
  • The prototype is powered by Genie 3, presented as a general-purpose world model that generates “the path ahead in real time” as the user moves and interacts, rather than rendering a static 3D snapshot [§Advancing world models].
  • The product exposes three capabilities: World Sketching (text + uploaded/generated images via Nano Banana Pro produce a preview image and let the user define perspective and character), World Exploration (real-time navigable environment with adjustable camera), and World Remixing (build on top of others’ prompts, browse a curated gallery, download videos of explorations) [§How Project Genie works].
  • Genie 3 is described as simulating physics and interactions for dynamic worlds, with “breakthrough consistency” claimed across long-form scenarios spanning robotics, animation, fiction, and historical settings [§Advancing world models].
  • Stated current limitations: outputs may not be true-to-life, may not adhere closely to prompts/images or to real-world physics; character controllability is inconsistent and can show higher latency; generations are capped at 60 seconds [§Building responsibly].
  • Promptable events (announced for Genie 3 in August 2025) are explicitly not yet exposed in this prototype [§Building responsibly].
  • Framed as supporting Google DeepMind’s AGI mission — world models are positioned as the ingredient that lets agents go beyond narrow-domain mastery (Chess, Go) toward “the diversity of the real world” [§Advancing world models].

Method

Section titled “Method”

Not disclosed. The post is a product announcement; the underlying Genie 3 model is described qualitatively (general-purpose, real-time path generation, physics + interaction simulation, multi-modal text+image conditioning, integrated with Nano Banana Pro for image preview and Gemini for unspecified roles). External coverage attributes an 11B-parameter autoregressive transformer running at 720p / 24 fps to Genie 3, but those numbers are not in the blog post itself.

The user-facing flow is: prompt the world (text + images), choose locomotion mode (walking, riding, flying, driving, …) and viewpoint (first/third person), define a character, get a World Sketching preview from Nano Banana Pro, then enter the world and navigate it. Worlds are bounded at 60s per generation. A remix path lets users build on top of prompts in a curated gallery.

Results

Section titled “Results”

No quantitative results. The post offers no benchmarks, no ablations, no comparisons to prior world models (DeepMind’s own Genie 2, NVIDIA Cosmos, Decart Mirage, open-source LingBot-World, etc.). It claims qualitative “breakthrough consistency” relative to static-snapshot exploration but does not substantiate the claim. The only concrete operating parameters mentioned are the 60-second generation cap and the eligibility envelope (US Google AI Ultra subscribers, 18+).

Why it’s interesting

Section titled “Why it’s interesting”

This is the closed-source/proprietary counterpart to the open Genie-3-class systems the wiki is already tracking — most directly Advancing Open-source World Models (LingBot-World) (LingBot-World), which explicitly positions itself against Genie 3 on the open/closed axis, and the NVIDIA Cosmos stack in World Foundation Models. For Luma the interesting bit is the product surface: World Sketching (image-grounded preview pre-commit), World Remixing (a gallery + remix layer that turns single-shot generations into a social object), and the locomotion/perspective conditioning UI. These are interface choices for a world model rather than model choices, and they suggest where the medium is heading even though the post discloses nothing about the model itself. The explicit not-yet-shipped feature — promptable events that mutate the world mid-exploration — is the obvious next axis to watch and overlaps with the action-conditioned-rollout work tracked under Autoregressive Video Generation (FlowAct-R1, Context Forcing, LongVie 2, LingBot-World).

See also

Section titled “See also”