PanoEvo keeps the world evolving in a panoramic latent: turn the camera away and back, and the action carries on
Synopsis
Video world models often freeze unobserved scenes when the camera turns away, so objects vanish or change appearance on return, because content outside the view has no representation that keeps changing. PanoEvo evolves the world state in the latent space of a panoramic video diffusion model, where every viewing direction is present at every time step, and a perspective renderer turns that state into sharp views. On WRBench it reaches the best overall average of 0.753 among 23 publicly evaluated models, versus 0.741 for the strongest baseline InSpatio-World 14B; on STEVO-Bench it scores 364.7 in total, versus 300.9 for ReMind.
Interpretation
When the camera turns away and returns, the scene no longer stays frozen at its last observed state; people and objects outside the view keep advancing. Earlier geometry caches, source-video re-rendering and memory mechanisms mainly retain what has been observed, and few densely represent unobserved content that keeps changing. On WRBench, PanoEvo attains an overall average of 0.753, above the 0.741 of the strongest baseline InSpatio-World 14B, with the best scores on visible state and both re-observation metrics.
A panoramic video diffusion model evolves the whole world state in latent space, and a perspective video diffusion model renders the requested view from it. The panoramic model contains every viewing direction at every time step, so the scene advances regardless of where the camera points; the perspective renderer turns the low-resolution panorama projection into sharp views. Training uses real 360-degree video: about 20K coarse clips for image-conditioned panoramic completion and 1,400 verified event clips for object-level state evolution; the renderer trains on 20,526 native-resolution clips.
The perspective renderer reads the panoramic state through latent-space projection and direction-restricted cross-attention, repairing the blur of direct projection. At a typical resolution a view covers only about 256 pixels horizontally and must be upsampled over three times, so direct projection blurs the view and degrades object structures. Compared with directly projecting the same panorama output, the renderer raises VBench imaging quality from 0.538 to 0.729 while largely preserving the evolved states.
Perspective
The demonstrated scope is camera rotation about a fixed center, with training and evaluation on fixed-camera clips built from real 360-degree video.
Camera translation brings parallax and newly visible surfaces that a single-center panorama cannot represent; local physical processes such as melting or liquid-level changes are rare in real 360-degree footage and underrepresented in the training data.
