Noiz, an AI company, and the Hong Kong University of Science and Technology (HKUST) have released Helix World, a model that turns a single image into a place you can move around in and generates that place’s sound at the same time. Press W, A, S or D to walk and the arrow keys to look. When the view turns, the audio shifts between the left and right ear with it. Noiz laid out the results on its own blog, alongside a research paper dated 29 September 2026.
Most models of this kind, called world models, produce silent video. Adding sound afterward does not fix that, Noiz argues, because a model that dubs a finished clip never learns where the camera went. A drum on your left stays put when you spin around. The team tested three such dubbing add-ons on a score for whether sound follows position on screen. They landed between 14.8 and negative 9.5, against 41.8 for Helix World. The test is Noiz’s own.
The speed claim is the first thing to read closely. The model runs at 24 frames per second and 768 by 512 pixels, on one NVIDIA H800, a data-center GPU. Its real-time factor is 0.77, meaning 0.77 seconds of computing for each second of video, decoding included. That is a number with named hardware behind it, which many launch posts skip. It also comes with limits the post states plainly: the paper reports no consumer graphics cards, and the picture is small. Output arrives in chunks of about 1.3 seconds, and a keypress only registers once the following chunk begins, so the world answers in steps rather than instantly.
Among the silent rivals in Noiz’s table, SANA-WM runs at 0.79 and Matrix-Game 3 at 0.58, so the speed is not unusual. Sound is the difference. Helix World scores 79.9 on a visual benchmark called WBench, behind the leading silent model at 82.0. With the sound muted, test viewers preferred a competitor, EchoWM, 44.7 percent to 26.3. With sound on, they preferred Helix World over EchoWM 67.9 percent to 20.8. The post does not say how many people took part.
How it keeps sound tied to position is explained only in outline. Helix World builds on LTX-2.3, an open-source model that makes video and audio together. Noiz feeds the camera path into the model’s attention over the picture, feeds keystrokes in separately, and says a link between the two streams “binds visual geometry to the audio features.” The post gives no account of how the model decides where a sound source sits, and the demonstrations are stereo only. Training data is the more concrete part. Noiz started with 4,100 hours of footage and kept 3,000 hours of true stereo, after discarding fake stereo with identical channels, narration and music beds.
Keeping the world stable while it streams is the other engineering problem. A teacher model that sees whole clips was distilled into a student that produces the stream one block at a time. Noiz says road and ground textures fell apart after 20 to 30 seconds under the standard method, and held together to 30 seconds with its added training step. Thirty seconds is the longest stretch the post shows.
The demonstration clips on the page, a fountain plaza, a sunlit street, a savannah ride and a pencil-sketch Paris, were picked by Noiz, and the previews are muted. Availability is the clear part. A live demo runs at helixworld.org, code is on GitHub, and a preview checkpoint is on Hugging Face. The post cites no outside evaluation. Even HelixBench, the 1,015-case audio-visual test the team released, was built by the people being scored on it.
Until someone outside Noiz runs the preview weights on a GPU a person can buy, 0.77 describes an H800 and nothing else.
Based on Noiz’s blog post introducing Helix World (undated), which cites a paper of 29 September 2026.
