Helix World: open weights turn one image into a playable video world
The LTX-based model streams explorable video with synced spatial audio from a single frame — Apache-2.0 weights are up, but real-time still wants an H800.
Copy markdown
Drive a video world with WASD
Start from one 768×512 image (plus an optional text prompt) and Helix World generates an explorable scene at 24 fps — WASD pans the camera, arrow keys change the view, and stereo audio pans along with it. The team reports stable ~30-second rollouts built in ~1.3-second chunks.
The weights are already on Hugging Face
NoizAI posted an Apache-2.0 preview checkpoint — a single safetensors bundling the transformer, audio/video autoencoder, and vocoder — next to the paper and a public GitHub repo. It is built on the open LTX-2.3 audio-visual model, distilled into a few-step causal student for streaming, so you can clone it and start poking this weekend.
The catch: it wants an H800
Real-time (RTF 0.77) is measured on a single NVIDIA H800, not consumer hardware. On WBench navigation it scores 79.9 against 82.0 for the best silent model, and audio sync is not class-leading — but on spatial-acoustic consistency it laps the prior audio world model EchoWM, 41.8 to ~12.7, and a user study preferred it 67.9% to 20.8%.