Wonder — Adobe's Real-Time Video World Model You Can Explore for Minutes
— by Vivax
Wonder, from Adobe Research and Johns Hopkins University (Jiacong Xu, Hanwen Jiang, Zhixin Shu, Kalyan Sunkavalli, Vishal M.
Adobe Research and Johns Hopkins University — Jiacong Xu, Hanwen Jiang, Zhixin Shu, Kalyan Sunkavalli, Vishal M. Patel and Yiqun Mei — have released Wonder, a general-purpose video world model for real-time, camera-controllable world exploration. From a single image or a conditioning video, Wonder builds a playable world you can navigate for up to a minute at 16 FPS with constant 0.5-second latency: move the camera, discover unseen regions, and return to previously observed areas to find them as you left them.
The paper's thesis is that interactive video world models suffer three chronic ailments — drifting camera controls, fading memories, and latency that grows with context — and that they must be fixed together, by system-level co-design. Wonder renders camera motion into a dense coordinate field whose renderings give the model spatially aligned motion and orientation cues, so camera control arrives as pixel-space visual evidence rather than abstract pose numbers. A sparse-attention memory keeps full-fidelity history KV caches and selectively retrieves only the relevant context tokens at inference time, holding latency constant no matter how long the rollout grows. And a rectified self-forcing distillation pipeline — a Mixture-of-Students design plus GAN Control Regularization — produces a real-time student that still respects the camera and keeps the teacher's diverse generation modes and long-term memory. Beyond image-to-video, Wonder supports video-conditioned generation, letting existing dynamic scenes be re-shot in real time.
For clinical world models the three ailments have direct analogues: control drift becomes unfaithful conditioning on interventions, fading memory becomes losing the distant patient history that explains today's state, and growing latency makes a model unusable at the bedside. Wonder's lesson — that control fidelity, long-horizon memory and constant-latency inference must be co-designed, not patched on — carries straight into medicine.