Helix World (preview)
Real-time interactive world model that generates video and spatial audio together.
Helix World, from Noiz AI and HKUST, generates an explorable video scene and matching spatial stereo audio from a single starting image (with an optional text prompt), steered in real time by camera motion and WASD/arrow-key input. It renders 768×512 at 24 fps in ~1.3-second chunks with a real-time factor of 0.77 on one H800 GPU, including decoding. Built on the open LTX-2.3 audio-visual model; the public checkpoint is a preview.
- Provider
- Noiz AI
- Type
- World model
- Released
- Oct 12, 2026
- Input
- text, image
- Output
- video, audio
- Open weights
- Yes
- License
- Apache-2.0
Best for
- Multimodal
- Research
- Real-time / low latency
Strengths
- Video + spatial stereo audio
- Faster than real time on one GPU
- Open weights