Noiz AI and The Hong Kong University of Science and Technology opened Helix World on OCTOBER 12: a real-time interactive world model that generates explorable VIDEO and camera-aligned STEREO SOUND together from a single starting image. You steer with WASD and the arrow keys; when the view turns, the SOUND FIELD turns with it. The team reports 24 frames per second on one NVIDIA H800 GPU, with a real-time factor of 0.77 at 768×512. That is NOT a laptop demo, and the public checkpoint is still labeled PREVIEW.
The project blog, arXiv paper (2609.38123, posted September 29), GitHub code, Hugging Face preview weights, and an interactive demo at helixworld.org are all live. For metaverse and XR builders, the POINT is less the scoreboard than the CLAIM: most interactive world models stay SILENT, and post-hoc video-to-audio pipelines do not see CAMERA MOTION, so they cannot PAN with the view.
Noiz AI and HKUST launched Helix World, a real-time world model that makes video and stereo sound together from one image. Steer with WASD; the sound turns with the camera. 24 fps on one NVIDIA H800. Code, preview weights out.
Video: Noiz AI
#NoizAI #HelixWorld #WorldModels— Hekate (@iCleanAI) October 13, 2026
What The Numbers Say
On the WBench navigation split, Helix World scores 79.9 while running FASTER than real time. Silent leaders such as Alaya-EVOKE-Turbo sit higher at 82.0 but need 4.29 seconds of compute per second of video. Among models with an RTF below 1.0, Helix World LEADS the table the authors publish, and it is the ONLY real-time entry there that also emits SOUND.
They also release HelixBench: 1,015 human-verified cases built to test whether stereo tracks on-screen sources under camera motion. Helix World’s spatial-acoustic consistency score is 41.8; EchoWM, the other audio-capable world model in the comparison, scores 12.7. Cascaded re-dubbing pipelines on the SAME video often score NEGATIVE on that spatial metric, which matches the intuition that a video-to-audio model never receives EGO-MOTION.
In a user study with audio, participants preferred Helix World over EchoWM 67.9% to 20.8%. With sound MUTED, EchoWM won that head-to-head, 44.7% to 26.3%. The ADVANTAGE the authors highlight is therefore MULTIMODAL, not pure picture quality.
How It Works, Briefly
Training data started from about 4,100 hours of raw footage (real-world, game recordings, and open-source sets), then a four-stage filter cut it to about 3,000 hours and 2.1 million clips with TRUE STEREO and metric CAMERA POSES. Dual-mono tracks, background music, and talking-head clips are the kinds of material the pipeline is designed to DROP.
The model builds on open-source LTX-2.3. A bidirectional TEACHER learns 6-DoF camera control and keyboard actions; a few-step causal STUDENT is distilled for STREAMING, with an online trajectory distillation loss aimed at long-horizon DRIFT. Generation runs in chunks of about 1.3 seconds, so a new key or prompt takes effect from the NEXT chunk, not INSTANTLY.
The Limits They List
Output stays at 768×512. Real-time speed is reported on a data-center H800, not a CONSUMER GPU. DeSync is not best-in-class on HelixBench, so onset timing still has room. And a PREVIEW weight is not a shipping PRODUCT for Quest, Vision Pro, or Steam Frame; it is RESEARCH tooling that XR teams can probe for spatial-audio WORLDS.
Sources: Noiz AI / HKUST blog; arXiv:2609.38123; NoizAI/HelixWorld; helixworld.org.