The Pulse
HelixWorld Preview Released Sept. 3; HelixBench Followed Oct. 12
Noiz AI and HKUST made HelixWorld’s preview code, checkpoint and browser demo available on September 3, 2026, before its technical report appeared on arXiv on September 29. Noiz’s October 12 post described the system and its HelixBench resu
AI.info Team · Written with AI ·
Noiz AI and the Hong Kong University of Science and Technology made HelixWorld’s Preview v1 inference code, checkpoint and browser demo available on September 3, 2026, according to the project repository. The technical report followed on arXiv on September 29. Noiz published a post introducing the system on October 12, the same date the repository lists for the release of HelixBench, a benchmark for interactive audio-visual world models.
HelixWorld generates an explorable video scene and stereo soundtrack together as a user moves through it. Starting from an image, users steer with WASD and turn the view with arrow keys; the model produces video at 24 frames per second on a single NVIDIA H800 GPU. Unlike systems that add sound after generating video, HelixWorld is designed to make the sound field shift with the camera.
When the view turns, the sound moves too
The central claim is not simply that the model can generate both picture and sound. It conditions both on camera motion and keyboard actions, so a sound source should pan as the viewpoint changes rather than remain fixed in the stereo field. Users can begin with an uploaded image, optionally add a text prompt, and redirect the scene as it runs. New controls take effect with the next generated chunk, which lasts about 1.3 seconds.
The paper’s authors describe the goal this way:
“We present HelixWorld, a real-time interactive audio-visual world model where visual scenes and camera-grounded spatial stereo sound co-evolve natively under user interaction.”
Lei Ke and co-authors, authors of “HelixWorld: A Real-time Interactive Audio-Visual World Model”
The September 29 paper says the system builds on an audio-visual model and uses a two-stage training process: a bidirectional model learns camera and action control, then a streaming version is distilled for lower-latency generation. The authors also assembled training footage with stereo sound and camera-pose information, filtering 4,100 hours of raw video down to 3,000 hours across 2.1 million clips.
One H800, 24 frames per second
Noiz reports a real-time factor of 0.77 on an H800 at 768×512 resolution and 24 frames per second, including video and audio decoding. A factor below 1 means the compute time is shorter than the duration of the generated output. On the WBench navigation split, HelixWorld scores 79.9; the reported comparison puts it behind some slower models but ahead of the other listed models that also run faster than real time.
HelixBench has 1,015 human-verified cases across six domains. The paper reports a spatial-acoustic consistency score of 41.8 for HelixWorld, compared with 12.6 for EchoWM, the other sound-generating world model in its comparison. The score measures whether stereo panning tracks the position of visible sound sources as the camera moves. HelixBench is the researchers’ own benchmark, so its results are not an independent evaluation.
Audio accounts for the clearest user-study advantage
In a blind pairwise study reported by the team, participants preferred HelixWorld with sound over EchoWM 67.9% to 20.8%; the remaining responses were ties. The system also beat versions of its own video redubbed with AudioX, ThinkSound and PrismAudio in the reported comparisons. With sound muted, however, HelixWorld lost to EchoWM, suggesting the reported preference advantage depends substantially on the generated audio rather than video quality alone.
The published benchmark results also show trade-offs: HelixWorld does not lead every audio measure. Other tested systems score better on temporal offset or text-to-audio alignment, according to the reported table. The project’s strongest case is specifically the connection between camera movement and spatial sound, not a universal lead across all audio or video measures.
A public demo, but demanding local hardware
The browser demo and preview checkpoint are available, and Noiz has published inference code on GitHub. The repository recommends an NVIDIA GPU with 80 GB of VRAM for local use and says a reviewed single-GPU run uses about 70 GB. The code carries an Apache 2.0 license; the checkpoint has a separate license.
The project’s public materials describe a research preview, not a general-purpose service with modest hardware requirements. For now, the browser demo lets users test whether a scene’s sound follows their movement. The published benchmark’s headline spatial score is 41.8 on HelixBench.