Helix World is a real-time audio-visual world model for creators and developers who want to explore AI-generated scenes with synchronized sound. Built by Noiz AI together with the Hong Kong University of Science and Technology (HKUST), Helix World turns a single starting image into a navigable environment whose picture and spatial stereo audio are generated together as you move, so the sound field follows your camera instead of being dubbed on afterward.
Core Features
- Generates video and matching stereo audio at 24 FPS from one image, conditioned on your camera motion and key presses.
- WASD to move and arrow keys to look; a new prompt or key press takes effect from the next roughly 1.3-second chunk.
- Runs on a single NVIDIA H800 at a real-time factor of 0.77, meaning it generates faster than real time including audio decoding.
- Built on the open LTX-2.3 audio-visual model; paper on arXiv, code on GitHub, and a preview checkpoint on Hugging Face.
- Scores 79.9 on the WBench navigation split and 41.8 on the team's own HelixBench spatial-acoustic consistency test, far ahead of EchoWM's 12.7.
Use Cases
- Game and virtual-world prototyping that starts from a single concept image instead of a 3D pipeline.
- VR and preview tooling where the sound field should rotate with the camera, not be dubbed on afterward.
- Research into audio-visual world models, ego-motion, and long-horizon generation without drift.
Pricing
Helix World is a free research preview. The interactive demo at helixworld.org, the Hugging Face checkpoint, and the GitHub code cost nothing, and there is no commercial API price yet. Running it yourself needs one H800-class GPU; there is no subscription.
Our Take
Best for researchers and indie creators who want navigable, sound-aware worlds without a render farm; the trade-off is that visual quality (WBench 79.9) still trails silent leaders like Alaya-EVOKE-Turbo (82.0), and this is a preview rather than a finished product. See the broader field in our AI Video tools category.
