Hugging Face has introduced Wonder, a novel video world model designed for real-time, camera-controllable world exploration. Given an image or a conditional video, Wonder constructs a playable environment where users can navigate interactively by moving the camera, discovering unseen regions, and revisiting previously observed areas over long horizons.
The system achieves this through a co-design of control method, memory mechanism, and training strategy. A key innovation is a camera conditioning approach using a dense coordinate field, whose renderings provide spatially aligned motion and orientation cues, allowing the model to interpret camera motion directly as visual evidence.
To handle long-term memory efficiently, Wonder employs a sparse attention-based memory mechanism that selectively attends to a small set of relevant context tokens at inference time, independent of the actual context length. This enables fast and precise memory retrieval over growing generation contexts.
The team also developed techniques to rectify the self-forcing-style distillation pipeline, improving the student model's ability to respect control signals while maintaining diverse generation modes and long-term memory from the teacher. These components enable Wonder to synthesize diverse, minute-scale videos at 16 FPS with coherent geometry, appearance, and dynamics across long rollouts.
Beyond image-to-video generation, Wonder naturally supports video-conditioned generation, allowing existing dynamic scenes to be re-shot in real time. The paper is available on Hugging Face Papers.