Solv3R Turns a Looped Transformer Into a Multi-View Geometry Solver With 16× Pipelined Throughput

taiyasaki · x · 2026-11-03

Solv3R turns a looped Transformer into a reusable learned multi-view geometric solver, letting geometric foundation models keep solving like classical SfM instead of fixed-set prediction.

One solver, every setting: handles offline batches, streaming frames, and arbitrary arrival order with the same logic, no task-specific training
What was solved stays solved: learns to preserve accurate states, remaining stable beyond its training horizon
New observations correct old ones: continues from current state and jointly refines old and new views
One cache for every call: solved views serve as reusable geometric context instead of per-layer/per-step KV caches
Pipelining: frames pipeline across refinement stages — 16 calls yield 16× solver throughput at the same latency

Original post →

Solv3R Turns a Looped Transformer Into a Multi-View Geometry Solver With 16× Pipelined Throughput

More from Multimodal