HelixWorld by Noiz AI: Free Demo & Setup

HelixWorld is Noiz AI’s free, open-source world model: walk through a scene made from one photo, with live video and stereo sound. See the demo and setup.

Most AI world models let you walk around a generated scene, but they’re silent. HelixWorld, from the voice AI company Noiz AI, makes the sound too, and the sound moves with you: turn left and the waterfall shifts to your right ear. This page explains what HelixWorld is, how to try the free browser demo, what it takes to run it yourself, what it costs, and what Noiz AI’s own paid plans include.

What is HelixWorld?

HelixWorld (Noiz also writes it as Helix World) is a real-time interactive audio-visual world model built by Noiz AI and The Hong Kong University of Science and Technology (HKUST). According to Noiz’s announcement post, you start from a single image, move with the WASD keys and look around with the arrow keys. As you move, the model generates the next stretch of video and the matching stereo sound at the same time.

• Video and sound together: the audio isn’t a soundtrack added afterwards. Picture and sound are generated in one pass, so a drum stops sounding when the drummer stops.
• Sound that follows the camera: the stereo field turns as the camera turns, which is what makes it feel like you’re inside the scene.
• Real time: it runs at 24 frames per second at 768×512 on a single NVIDIA H800 data-center GPU, slightly faster than real time.
• Steer it live: generation runs in chunks of about 1.3 seconds, and each key press or new text prompt takes effect from the next chunk.
• Any starting picture: photos, game screenshots or drawings. A pencil-sketch street stays a pencil sketch as you walk through it.

One thing to be clear about: despite the “3D world” label you’ll see online, HelixWorld doesn’t build a 3D model or mesh you can export. It generates video frame by frame in response to your movement, the way other interactive world models do. It’s built on LTX-2.3, an open-source audio-visual video model.

Release timeline

• September 3, 2026: Noiz released the HelixWorld Preview v1 inference code on GitHub and a preview checkpoint on Hugging Face, plus the browser demo.
• September 29, 2026: the technical report, “HelixWorld: A Real-time Interactive Audio-Visual World Model,” went up on arXiv (2609.38123).
• October 12, 2026: Noiz published its “Meet Helix World” post and released HelixBench, a test set of 1,015 human-checked cases that measures whether a model’s sound follows the camera.

The public model is still labelled a preview. Noiz says so itself in the post’s limitations section.

Is HelixWorld free?

Yes. The browser demo at helixworld.org is free to open: you choose with or without sound, then pick one of the sample worlds or upload your own photo. The HelixWorld code on GitHub is Apache-2.0, and the preview weights on Hugging Face are published under Apache-2.0 too. Two parts come with other terms: the bundled LTX-2.3 code from Lightricks falls under the LTX-2 Community License, and the Gemma text encoder under Google’s own terms. Read those before any commercial use.

HelixWorld Pricing (2026)

HelixWorld itself has no price. What it costs depends on how you use it:

Option	Cost	What you need
Browser demo	Free	A web browser
Code and weights	Free (Apache-2.0)	A GitHub and Hugging Face download
Run it yourself	Your own GPU or cloud GPU rental	Linux, an NVIDIA GPU with about 80 GB of VRAM
Noiz AI plans	From $6 a month	Only for Noiz’s voice, music, image and video tools, not for HelixWorld

The real cost of self-hosting is the GPU. Noiz’s reviewed single-GPU run in BF16 uses about 70 GB of VRAM, so you need an 80 GB data-center card such as an H100 or H800, either your own or rented by the hour. Noiz hasn’t published results on consumer graphics cards.

How to try HelixWorld

In the browser

1. Open helixworld.org and choose “enter with sound” (use headphones to hear the stereo effect) or “enter without sound.”
2. Pick a sample world, or click Upload your photo. Your image is encoded at 768×512.
3. Add a text prompt if you want. If you leave it empty, the scene is described automatically from the image.
4. Press Enter, then move with W, A, S, D and look around with the arrow keys. Type a new prompt to change the scene as you go.

On your own GPU

You need Linux, Python 3.11, CUDA 12.x, an NVIDIA GPU (80 GB of VRAM recommended), and system ffmpeg and ffprobe. The setup from the HelixWorld README looks like this:

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

The download script fetches the HelixWorld weights and a Gemma 3 12B text encoder. The --actions string is your path through the world: W A S D move, left, right, up and down turn the camera, and you can combine keys (W+D:8). The finished clip is saved as an MP4 under the output folder.

What is Noiz AI?

Noiz AI is the company behind HelixWorld, and most people searching “noiz ai” are looking for its main product: an AI voice and audio platform. It offers emotional text-to-speech, voice cloning, voice design, sound design, music creation (its YuE2 model turns ideas into songs), AI image and video generation, and Noiz Studio, an all-in-one audio and video workspace with 100+ templates for things like music videos and video translation.

Noiz AI pricing

These plans are for Noiz’s creative tools, not for HelixWorld. From the Noiz pricing page on October 18, 2026:

Plan	Monthly billing	Yearly billing	Credits per month	Highlights
Lite	$6/month	$4.50/month	100,000 (about 100 minutes of audio)	Voice cloning, voice design, music, 1,000 characters per generation
Pro	$19/month ($9.50 for the first month)	$14.25/month	300,000 (about 300 minutes of audio)	10,000 characters per generation, watermark-free exports, commercial use
Ultra	$50/month ($24.99 for the first month)	$25/month	1,000,000 (about 1,000 minutes of audio)	20,000 characters per generation, watermark-free exports, commercial use

If you mainly need AI voices, compare Noiz with ElevenLabs, the best-known option, before you subscribe.

Limitations

• Preview model: the public checkpoint is a preview release.
• Low resolution: output is 768×512.
• Data-center GPU needed: real-time speed is measured on an NVIDIA H800, and local runs need about 70 GB of VRAM.
• Not the sharpest picture: on visual quality alone it’s slightly behind the best silent world models, and with sound muted viewers preferred EchoWM. Its edge is the audio.
• No 3D export: it makes video, not a 3D scene file you can open in a game engine.

HelixWorld alternatives

• EchoWM: the only other world model in Noiz’s comparison that also makes sound, but it runs slower than real time.
• Silent world models: Matrix-Game 3, SANA-WM, LingBot-World v2 and HY-WorldPlay 1.5 focus on picture and control only.
• If you want to build an actual game: world models are still research demos. See our guide on how to make a game with AI for tools that output playable games.
• If you just want AI video clips: a regular video model is simpler. Start with our list of free AI text to video generators.

Verdict

HelixWorld is one of the most fun AI demos to try right now: upload a photo, put on headphones and walk into it, with sound that turns as you do. It’s free, open source under Apache-2.0, and honest about being a preview. For most people the browser demo is the way to use it, because running it locally takes an 80 GB GPU. If you came here for Noiz AI’s voice tools, its paid plans start at $6 a month.

HelixWorld adds what other world models miss: sound that moves with the camera. The free demo makes it easy to try, and the Apache-2.0 code and weights invite experiments.

Treat it as a research preview, not a production tool. Use the browser demo unless you have an 80 GB GPU.