Open source (Apache 2.0)
$0Self-hostedPreview v1 inference code and checkpoint. Hardware is your own: an NVIDIA GPU with 80 GB VRAM is recommended, and a BF16 single-GPU run uses about 70 GB. Official source 1 ↗ · Official source 2 ↗
Reviewed tool profile
HelixWorld
Open-source real-time world model from Noiz AI that turns one image into an explorable scene with camera-aligned spatial sound.
Sources checked
- Pricing
- The inference code and the Preview v1 checkpoint are both released under the Apache 2.0 license. Running it yourself needs your own NVIDIA GPU (80 GB VRAM recommended). Official source 1 ↗ · Official source 2 ↗
- Available as
- Browser demo at helixworld.org, controlled with WASD and arrow keys or by dragging; on phones it asks for landscape orientation. · Self-hosted inference on Linux with Python 3.11, CUDA 12.x, an NVIDIA GPU and system ffmpeg/ffprobe. Official source 1 ↗ Official source 1 ↗
- Primary group
- Video & Animation
What it does
HelixWorld is a real-time interactive audio-visual world model from Noiz AI. Start from one image and an optional prompt, then move through the generated scene while picture and camera-aligned stereo sound are generated together. It ships as a browser preview demo and as open-source Preview v1 inference code and weights.
Official source 1 ↗Features
- World from a single image. Pick a sample frame or upload any photo; it is encoded once at 768 × 512 and becomes the world you enter. Official source 1 ↗
- Automatic or custom prompt. The prompt is optional: an empty prompt is written from the starting frame, and your own words are kept and refined. Official source 1 ↗
- Live keyboard navigation. The world is generated live in 1.3-second chunks; WASD moves and the arrow keys look around. Official source 1 ↗
- Change the scene mid-session. A new prompt can be typed at any time; the current world keeps playing and the change applies from the next chunk. Official source 1 ↗
- Camera-aligned spatial stereo sound. Sound is generated together with the picture rather than dubbed afterwards, and the stereo field turns with the camera. Official source 1 ↗ · Official source 2 ↗
- Real time on one GPU. Noiz AI reports a steady-state real-time factor of 0.77 at 768×512 and 24 FPS on a single NVIDIA H800, including video and audio decoding. Official source 1 ↗
- Scripted MP4 rendering. The open-source runner takes a first frame, prompts and an action script (W, A, S, D, left, right, up, down, stop) and writes a clean MP4. Official source 1 ↗
Pricing & free limits
Billing details
Billing details were not established by the checked official sources.
Free access & limits
Free access and its limits were not established by the checked official sources.
Best for & limitations
Turning a photo into an explorable scene with sound, directly in the browser. Official source 1 ↗ · Official source 2 ↗
Rendering first-person navigation clips with synchronized audio from a scripted camera path on your own GPU. Official source 1 ↗
Research on audio-visual world models, using the companion HelixBench benchmark for spatial-acoustic consistency. Official source 1 ↗
- Preview release. What is available is Preview v1: inference code and a preview checkpoint, not a finished product. Official source 1 ↗
- Needs a data-centre GPU to self-host. Self-hosting is Linux-only with CUDA, and a single-GPU BF16 run uses about 70 GB of VRAM. Official source 1 ↗
- Slight input delay. Controls apply to the next generated chunk, so the scene responds a moment after you press a key. Official source 1 ↗
- 768 × 512 output. The starting frame is encoded at 768 × 512, which sets the generated resolution. Official source 1 ↗ · Official source 2 ↗
Similar tools
Tools like HelixWorld
Published tools that share its categories.
ngramAI video maker for business teams: onboarding, training, demos and explainers from a prompt, screen recording or doc, edited by chatting.
Visit 
ReelsyAI product video and image studio for small ecommerce teams, with multi-model video generation, a chat-based Video Agent and a Codex/ChatGPT plugin.
Visit 
ArtCraftOpen-source AI image and video studio where you stage shots on a 2D canvas or in 3D, then let models like Nano Banana and Seedance render them.
Visit 
GeminiGoogle's AI assistant for chat, document summaries, Deep Research reports, writing, coding, and image, video and music generation.
Visit 
MiniMax H3MiniMax’s multimodal video model: 2K clips up to 15 seconds with native stereo sound.
Visit 
Vidu Q4Vidu’s preview video model that turns images or references into up to 16-second, 4K clips with native audio.
Visit