Helix World generates video and spatial audio in one pass

Helix World, a preview world model from Noiz AI and the Hong Kong University of Science and Technology, generates video and matching spatial stereo audio in the same pass and lets you steer the camera while it runs. The announcement is dated October 12, 2026, and the demo is open at helixworld.org.

The announcement page describes the interaction plainly: you start from a single image, move with WASD, look around with the arrow keys, and the model generates the next stretch of picture and sound together. A text prompt is optional, and if you leave it empty the scene is described from the frame. Generation runs in chunks of about 1.3 seconds, and each keyboard input takes effect from the next chunk. When the camera turns, the page says, "the sound field turns with it."

The maker's own material shows continuous generation across eight named scenes, listed on the GitHub README as Savannah Ride, Crystal Cave, Sketch Studio, Vineyard at Dusk, Pencil Paris, Country Lane, Sunlit Street and Fountain Plaza. One highlight describes a drumming clip with a visible pause: HelixWorld's transients subside during the pause and resume with the motion, while a post-hoc dubbing baseline, AudioX, produces a transient during the pause. Those are the makers' clips, and this article reports what their text says the clips show.

The technical report gives the numbers. Helix World builds on the open-source LTX-2.3 audio-visual model, adds 6-DoF camera control through PRoPE and keyboard actions through AdaLN-Zero, then distills the result into a causal student. It reaches a steady-state real-time factor of 0.77 at 768x512 and 24 frames per second on a single NVIDIA H800, including video and audio decoding. Output tops out at 768x512.

The team's own benchmarks put the sound first. On HelixBench, the benchmark the team released with the model, Helix World scores 41.8 on spatial-acoustic consistency against 12.7 for EchoWM, the only other world model in the comparison that generates sound. The picture is less dominant: on the WBench navigation split Helix World scores 79.9, behind silent models including Alaya-EVOKE-Turbo at 82.0. A write-up at Grainzilla reports that in the team's user study viewers preferred Helix World over EchoWM 67.9% to 20.8% with audio on, but with the sound muted they preferred EchoWM 44.7% to 26.3%.

That points at what it means for creators. This is previs and mood-board territory, not a source of final shots. Walking a location from a single still and hearing how the space would sound could help plan ambience, blocking or a game-style walkthrough. The catch is hardware. The repository recommends an NVIDIA GPU with 80 GB of VRAM and notes that a reviewed BF16 single-GPU run used about 70 GB, which puts local use out of reach for most editing machines. The browser demo is the realistic way to try it.

Code is Apache 2.0, and the weights carry their own license published with the preview checkpoint; the model card is tagged Apache 2.0 and points to its own LICENSE file.

Try Helix World in the browser at helixworld.org, or pull the preview checkpoint from Hugging Face under its own weights license.