Game AI Daily · Thursday, November 19, 2026
GAMEAI HOT
DeepMind opens Genie 4 world model to Google AI Ultra subscribers
Google DeepMind has rolled out a new version of Project Genie to eligible Google AI Ultra subscribers, opening up the Genie 4 world model for the first time; users can enter prompts during exploration to change the weather, add characters or alter the story direction in real time, and it is paired for the first time with real-time native audio that changes with the scene.
Model releases & updates
Inworld releases SAVA voice activity and turn detection model
Inworld released SAVA, a semantic acoustic voice activity model that merges voice activity detection and turn detection into a single model and is already running in its voice products. The model outputs two probabilities every 40 ms, for speech and end of turn, has 25.5M parameters, and processes each 40 ms frame in under 3 ms on a single CPU thread of an Apple M3 Max.
fal releases image-to-3D models Idol Max and Idol Max Turbo
fal released two image-to-3D models, Idol Max and Idol Max Turbo, built and optimized in-house on top of Idol. A single input image produces game-ready 3D assets: the output is a real mesh in GLB format rather than a Gaussian splat, with UVs and 2K textures, plus optional PBR materials and automatic rigging.
Inworld open-sources voice isolation model Voice Isolator with offline and real-time versions
Inworld released Voice Isolator, its first open-source model, which removes background voices, music, and transient noise while preserving the speaker's timbre, breathing, and volume. It works at 48 kHz and comes in two checkpoints: offline and real-time streaming with 20 ms lookahead.
Engines & tools
NVIDIA open-sources MotionBricks, generating character motion in real time from 350,000+ motion clips
NVIDIA has open-sourced MotionBricks, a model trained on 350,000+ motion clips that generates character motion in real time from target poses. The release includes the pretrained model, creation tools, and a reference implementation built with Newton; the code is on GitHub (NVlabs/motionbricks).
Unreal Engine livestream breaks down MetaHuman Animator markerless motion capture plugin
Unreal Engine's Inside Unreal livestream took a deep dive into the new MetaHuman Animator Markerless Motion Capture plugin, which turns ordinary video shot on a phone or webcam into full-body animation for MetaHuman characters.
Snap introduces Easy Lens CLI to create Snapchat Lenses with AI coding assistant
Snap has introduced Easy Lens CLI, letting developers build Snapchat Lenses by describing requirements in natural language within AI coding assistants such as Claude Code, Codex, and Cursor. The tool is now live and free, and requires a Snapchat account.
Games & apps
Epic AI MegaJam 2026 wraps up: 10,000 sign-ups and over 3,000 games submitted in 4 days
Epic-hosted AI-themed game jam "Epic AI MegaJam 2026" received more than 3,000 valid entries during its 4-day development window from November 12 to 15, with registrations exceeding 10,000.
Voicemod brings real-time voice skins directly into League of Legends
Voicemod announced a partnership with Riot Games, and eligible players in North America and Oceania can now use real-time voice skins in League of Legends voice chat on PC and Mac without downloading the separate Voicemod app.
Will Wright's new simulation game MyMems will use AI to mine content from players' memories
Will Wright's Gallium Studios has announced MyMems, a new simulation game that uses generative AI plus a self-developed behavior engine to turn memories entered by players into simulated scenes and characters. The project, formerly Proxi announced in 2018, was halted in late 2024 with layoffs after running out of funding, and is now revived through a partnership with AI middleware company Artificial Agency.
Research
InertialWorld: A video world model with inertia as a prior
Researchers propose InertialWorld, which treats inertia as a structural prior for autoregressive video prediction, letting existing scene evidence carry over time instead of being rebuilt at every step. Its core Inertial Memory maintains both a dense Eulerian grid and persistent Lagrangian carriers, transports carrier content according to action-conditioned motion, and then fuses old and new content with a learned gate. In experiments on robotic manipulation, 3D game simulation, and open-world navigation, the model's autoregressive prediction performs comparably to strong video world model baselines, and the code is open-sourced on GitHub.
Towards data-efficient and real-time world models
The paper proposes replacing the next-token prediction paradigm with a two-stage discrete diffusion method and introduces a lightweight refinement Transformer for iterative masked prediction, achieving a 79x model speedup over the autoregressive world model of Kanervisto et al. 2025 with only a small quality loss. The authors also show that the model can be trained on about one day (28 hours) of gameplay data to produce high-quality, coherent gameplay sequences, pointing to the usability of world models in settings constrained by real-time inference and data efficiency.
Briefs
- EyeHeight benchmark reveals camera-height regression in interactive video world modelsarXiv Game World Models & Generation
- Steam indie game Zombies Don't Exist goes viral: real city maps and tough choicesGameLook