Updated

Helix World (preview)

Noiz AI and HKUST introduced Helix World, a real-time interactive world model that generates video and spatial stereo audio together from a starting image, conditioned on camera movement and keyboard input. A preview checkpoint is available on Hugging Face alongside a public demo, paper, and code.

By No Code MBA · Facts checked

Provider
Noiz AI
Announcement date
Status at last check
Limited access or preview
Model family
Helix World
Capabilities
Video, Audio, Image

What changed

Helix World, from Noiz AI and HKUST, is a real-time interactive world model that turns a single starting image into explorable video with stereo spatial audio that pans with camera motion. Users move with WASD and look with arrow keys. It is built on the open LTX-2.3 audio-visual model, distilled into a few-step causal model, and outputs 768x512 video at 24 fps in roughly 1.3 second chunks, with a reported real-time factor of 0.77 on one H800 GPU.

Access and availability

A preview checkpoint is on Hugging Face at NoizAI/HelixWorld-preview (listed as Apache 2.0, with full terms in the repo LICENSE), code is on GitHub at NoizAI/HelixWorld, and a public demo runs at helixworld.org.

Official access documentation

What you can build

Try the preview on a small set of starting images and camera movements, then review whether generated motion and stereo sound remain coherent as the viewpoint changes.

Limitations

Noiz AI calls the public checkpoint a preview release. Output is capped at 768x512, visual quality trails the best silent world models (79.9 vs 82.0 WBench average, per Noiz AI), and reported speeds are on a data-center H800 GPU, not consumer hardware.

Sources and release history