Noiz AI and HKUST Release Helix World, a Video and Stereo Sound World Model
Summary
Noiz AI and HKUST release Helix World, an interactive world model that generates video and stereo sound together from a single image, steered with WASD and arrow keys. The developers report 24 frames per second at 768x512 on one NVIDIA H800 and a 79.9 WBench navigation score. A preview checkpoint is on Hugging Face.
Key Points
- Noiz AI and HKUST report a HelixBench spatial-acoustic consistency score of 41.8, against 12.7 for EchoWM, the other sound-generating model.
- A user study finds participants prefer Helix World over EchoWM 67.9 percent to 20.8 percent on full audio-visual rollouts.
- Training draws on 3,000 hours of stereo clips curated from 4,100 hours of raw footage.