Adobe Research and JHU debut Wonder, a 16 FPS video world model

A new report from Adobe Research and Johns Hopkins, posted to Hugging Face's papers hub, describes a video world model called Wonder that lets a user steer a virtual camera through a scene generated from a single image or a conditioning video, at 16 frames per second, over what the authors call minute-scale rollouts. That combination, camera control, persistent memory, and real-time output at the same time, is the part the paper is arguing has not been done well before.

The interesting design choice is how the camera is fed to the model. Instead of handing the diffusion model an abstract pose or a Plücker-ray encoding and asking it to work out what that means for pixels, the authors render a synthetic 3D scaffold and environment map along the target camera trajectory and feed those renderings in as visual evidence. Translation, rotation, and parallax show up as frame-aligned cues the model can copy. Paired with that is a sparse attention memory that only attends to a small set of relevant historical tokens at each step, which is how the authors keep per-step latency roughly flat as the rollout grows and how the model can revisit an area it already generated without the scene mutating underneath it.

On the training side, Wonder follows a teacher-to-student distillation pattern. A bidirectional many-step video diffusion teacher is converted into a causal autoregressive generator and then distilled into a few-step streaming student using a modified self-forcing pipeline. The authors are candid that self-forcing style distillation tends to collapse modes and degrade fine details, and their answer is a timestep-wise mixture of students, three separate generators covering coarse structure, refinement, and detail across a four-step sampler.

The honest caveat is that this is a research report, not a product. There is no release date, no licensing statement, no parameter count, no inference-hardware disclosure, and the abstract-level material does not put Wonder head-to-head against named competitors with benchmark tables you can read in one sitting. Take the specifics as claimed by the authors, not as independently verified.

What is worth watching is where this lands inside Adobe. A camera-controllable, minute-long, interactive video generator is the natural next step for previsualization, virtual production, and interactive storyboarding, and it is the class of capability that turns a research paper into a Creative Cloud feature. The startups whose main pitch was 'real-time interactive world model' now have a well-resourced incumbent making the same pitch.