Adobe Research and Johns Hopkins are transforming an image or video into a navigable environment at 16 frames per second, without any control drift or memory loss.
Presented at the end of July 2026 by Adobe Research and Johns Hopkins University, Wonder addresses three recurring flaws in video world models: the progressive drift of camera controls, the erasure of memory of previously visited scenes, and the latency that swells as the context grows longer. The proposal is to deal with these three problems jointly rather than separately.
The camera movement is no longer transmitted as a simple abstract geometry but rendered as visual clues in the pixels space, making control much more faithful. For memory, the model keeps the complete history in full resolution in its caches and recovers only relevant fragments out of careful attention: latency remains constant regardless of the duration of the exploration. A distillation phase finally produces a student model capable of turning in real time while continuing to obey the camera.
Result: from a fixed image or video, a user can move his point of view into the scene generated at 16 images per second, on interactive sequences of the order of the minute — a horizon so far out of reach for this type of system.
Read the full article on the original source
Go to the source →