Introduction

Most world models let you walk through a generated scene in total silence. Helix World, from Noiz AI and HKUST, makes the sound part of the world. Noiz officially launched it on October 12, 2026: start from a single image, move with WASD, look around with the arrow keys, and the model generates the next stretch of video and spatial stereo audio together at 24 fps. Turn the camera and the sound field turns with it.

You can try it today in the browser at helixworld.org. The inference code is on GitHub, and a Helix World Preview v1 checkpoint is on Hugging Face. The team also released HelixBench today, a new benchmark that checks whether a world model's stereo actually tracks the camera.

What Shipped

  • Model: Helix World, an interactive audio-visual world model (image in, explorable video plus stereo out)
  • Who: Noiz AI with The Hong Kong University of Science and Technology
  • Resolution and speed: 768×512 at 24 fps, generated in chunks of about 1.3 seconds
  • Real-time factor: 0.77 on one NVIDIA H800, including video and audio decoding (below 1.0 means faster than real time)
  • Controls: WASD to move, arrow keys to look, plus an optional text prompt you can change mid-session
  • Open bits: code on GitHub (Apache 2.0) and a preview checkpoint on Hugging Face
  • New today: HelixBench, with 1,015 human-verified cases across six domains

Why Sound Changes the Game

Sound tells you things the picture can't. You hear what a surface is made of, how hard two things hit, and what's going on off screen. Adding audio afterward with a video-to-audio model doesn't fix this, because that model never sees where the camera is going. Its sound ignores your movement and adds latency on top.

Helix World generates both streams in one pass. The camera path and your keyboard actions feed straight into the model, and cross-modal attention ties the visual geometry to the audio. That's how a fountain on your left slides into your right ear as you turn around.

The Numbers

On the WBench navigation split, Helix World scores 79.9. That puts it close to the best silent world models while running faster than real time. Among the faster-than-real-time models in the paper's comparison, it scores highest: SANA-WM gets 76.0 and Matrix-Game 3 gets 71.3. The three models that score higher (Alaya-EVOKE-Turbo at 82.0, EchoWM and Zing-0.5 at 81.0) all need between 1.98 and 4.29 seconds of compute per second of video.

On sound, the gap is wider. Helix World's spatial-acoustic score on HelixBench is 41.8. EchoWM, the only other world model in the comparison with native audio, scores 12.6, and re-dubbing Helix World's own video with AudioX, ThinkSound or PrismAudio scores between 14.8 and minus 9.5. In the user study, people preferred Helix World's full audio-visual rollouts over EchoWM by 67.9% to 20.8%.

The team is upfront about where it trails. With the sound muted, viewers preferred EchoWM's picture, and LTX-2.3 and the ThinkSound pipeline have tighter onset timing.

How It Was Built

  • Base model: the open-source LTX-2.3 audio-visual model from Lightricks
  • Data: 4,100 hours of raw footage filtered down to 3,000 hours of true stereo (2.1 million clips, 48 kHz) with metric camera poses for every frame, rejecting background music, narration and talking heads
  • Control: 6-DoF camera paths enter through PRoPE positional encoding, and keyboard actions enter through AdaLN-Zero
  • Streaming: a bidirectional teacher is distilled into a few-step causal student with a sliding KV cache, then trained with self-forcing plus an online trajectory-distillation loss so 30-second rollouts stay coherent instead of melting into mush

Run It Today

The easy path is the Helix World demo. Pick a starting image or upload your own, leave the prompt empty if you like (the scene gets described for you), then steer. A sketch-style input stays sketch-style, so stylized art is fair game.

To run it locally, you'll need serious hardware: Linux, Python 3.11, CUDA 12.x, ffmpeg, and an NVIDIA GPU with 80 GB of VRAM recommended. The README says a BF16 single-GPU run uses about 70 GB.

git clone https://github.com/NoizAI/HelixWorld.git
cd HelixWorld
conda create -n helixworld-preview python=3.11 -y
conda activate helixworld-preview
python -m pip install -r requirements.txt
python download_models.py
CUDA_VISIBLE_DEVICES=0 ./run.sh \
  --image /path/to/first_frame.png \
  --prompt-file examples/prompt.json \
  --actions "W:5,right:5,stop:5" \
  --perspective first_person \
  --num-frames 121 \
  --output-dir outputs/demo

Actions are W A S D left right up down and stop, and you can combine them with a plus sign, like W+D:8. The number after the colon counts latent transitions, and a 121-frame clip has 15.

One note on licensing: the code is Apache 2.0 and the checkpoint card is tagged Apache 2.0, but the card says its own LICENSE file sets the terms, and the model builds on LTX-2.3. Read that file before you use it commercially.

Limitations

  • Output tops out at 768×512
  • Real-time speed is measured on an H800 data-center GPU, and the paper doesn't report consumer cards
  • The public checkpoint is a preview release
  • On picture quality alone it sits slightly behind the best silent models (79.9 against 82.0 on WBench)

Original Source

Conclusion

Silent world models were always a little eerie, like walking through a dream with your ears plugged. Helix World is a real step toward generated places you can hear as well as see, and it's running in a browser tab right now. Go turn your head next to that fountain.

—Aurelia ♡