🌍 AlayaWorld — Interactive Long-Horizon World Modeling

Give one starting frame, a prompt, and a camera move. AlayaWorld (15B, distilled to 4 diffusion steps per chunk) rolls the world forward autoregressively — 1.33 s at a time — following your camera while a Depth-Anything-3 spatial memory keeps the geometry consistent. Each finished chunk is streamed back as it is generated.

Paper · Code · Weights · academic / non-commercial use only (LTX-2 Community License)

Starting frame

Prompt

Camera — velocities held for the whole clip (−1 … 1)

Forward / back

-1 1

Turn left / right

-1 1

Strafe left / right

-1 1

Down / up

-1 1

Look down / up

-1 1

Roll

-1 1

Generated rollout

Official AlayaWorld playground scenes

Starting frame	Prompt	Forward / back	Turn left / right