Figure 3: ARROW architecture. The multi-view encoder E tokenizes each frame independently
and employs global attention to produce a shared global context F together with an ID token h for
i
each frame. The query encoder combines a source-centered RGB crop, the Fourier-encoded (u, v)
coordinate, and the ID tokens of the queried source, target, and reference view frames into a latent
query token z . Each latent query token independently cross-attends to the global context, so a single
q
encoding serves arbitrarily many queries and yields a consistent 4D representation of the scene.