Noiz AI and HKUST open-source Helix World, a real-time audio-visual world model
lmoroney · x · 2026-10-17
Noiz AI and HKUST open-sourced Helix World, a real-time interactive audio-visual world model: start from one image (plus optional text), move with WASD and look with arrow keys, and it generates the next chunk of video together with matching 48 kHz stereo audio — the sound field turns with your head.
- 24 fps at 768x512 on a single NVIDIA H800, real-time factor 0.77 including decoding; generation runs in 1.3s chunks
- WBench navigation score 79.9, close to stronger silent world models; on the newly released HelixBench spatial-acoustic metric it scores 41.8 vs 12.7 for EchoWM, the only other sound-generating world model compared
- Code on GitHub, preview checkpoint on Hugging Face, paper on arXiv, live interactive demo at helixworld.org
- Stated limits: 768x512 resolution, preview weights, real-time numbers are H800-specific
More from Multimodal
- Voice-prompting industrial techno: describing clacking keyboard sounds to an AI music model — michellechen · 2026-10-17
- UniSlider adds a perceptual slider to control how much an AI image edit changes — CSProfKGD · 2026-10-17
- MiniMax-H3 plus ORB360 LoRA demoed making comics with freely movable camera angles — MiniMax_AI · 2026-10-17
- You Can Generate 3D HTML Animations Directly in Runway's Agent Mode — CurieuxExplorer · 2026-10-17
- Open-source ComfyUI node fixes MOSS-Image's resolution-dependent BBox coordinates — b4ldur · 2026-10-17
- AI video pipeline revealed: Blender blockout to Seedance with three generative models — apostraphi · 2026-10-17