Noesion Radarnoesion.ai
Release2026-09-29

Noiz AI and HKUST introduced Helix World. It generates explorable video and stereo sound together at 24 fps on one GPU. Users steer with WASD and arrow keys. Sound pans with camera motion. It scores 79.9 on WBench and 41.8 on HelixBench for spatial audio. The demo, paper, code and weights are public. Training used 3,000 hours of curated stereo video. A bidirectional teacher model was distilled into a causal streaming model. Progressive training reduced camera and translation errors. The public release is a preview checkpoint.

What changed

World models generated silent video or added sound after the fact with latency and misalignment.

What it unlocks

Exploring a generated scene with camera-aligned stereo sound in real time.

  • 24 fps on one NVIDIA H800 GPU
  • 79.9 WBench navigation score
  • 0.77 real-time factor
  • 41.8 HelixBench spatial-acoustic consistency

What you need to act on it

  • an NVIDIA H800 GPU
  • a starting image

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.