Solv3R Turns a Looped Transformer Into a Reusable Multi-View Geometry Solver

taiyasaki · x · 2026-11-03

Solv3R, from UBC and Sony researchers, trains a looped Transformer to act as a learned multi-view geometry solver.

Unlike geometric foundation models that predict all geometry in one forward pass, Solv3R keeps solving iteratively like classical SfM; its update operator is independent of iteration count, so refined and newly added views can be jointly updated
Training preserves accurate reconstructions under continued iteration and corrects inaccurate states, enabling both refinement of fixed image collections and incremental integration of new observations
Representations of solved views can be cached once and reused as geometric context across all solver calls, giving compact key-value caching
One model, without task-specific training or modules, supports offline reconstruction, incremental reconstruction, relocalization, and streaming frame-by-frame

More from Embodied