Fig. 4: Method Overview. Our method performs streaming 3D reconstruction via
dual-branch inference. At time step t, the main branch predicts depth and pose from
state s using both image and RayMap features. The predicted pose Tˆ is remapped
t−1 t
into RayMap and encoded into tokens r′ , then queried against the frozen state s
t t−1
by the RayMap branch to obtain a static-biased prediction. The depth difference
between branches is pro jected via image-state attention to form staticness weights α
t
that suppress dynamic regions during state update (s = s + α ⊙ ∆s ). This dual-
t t−1 t t
branch scheme identifies and suppresses dynamic regions at inference time.