Key Features
Generates interactive video worlds from images or source videos.
Supports camera-controlled navigation through generated environments.
Maintains full-fidelity history through a memory mechanism and sparse attention.
Uses dense control signals for translation and rotation guidance.
Targets coherent geometry, appearance, and dynamics across long rollouts.
Demonstrates game, cartoon, and real-world generation examples.
Supports video-conditioned generation as well as image-to-video worlds.
Links to an arXiv paper for technical details and evaluation.
About Wonder
The system combines dense control signals, a full-fidelity memory mechanism, sparse attention over history, and a rectified distillation training strategy. Its design aims to maintain geometry, appearance, dynamics, and interactive control across real, cartoon, and game-like worlds.
Wonder is useful for researchers building controllable video world models, interactive scene generation systems, and game-like AI environments. It is especially relevant when constant-latency navigation and long-horizon scene consistency matter more than one-off video sampling.