ZipMap zips 3D reconstruction to linear time

ZipMap is interesting because it does more than speed up a benchmark: it reframes large-scene reconstruction as stateful memory compression instead of ever-growing global attention. That makes it one of the clearest signs that test-time training can be useful outside language modeling.

• –The core trick is swapping quadratic global attention for local window attention plus large-chunk test-time training layers, pushing runtime from O(N²) to O(N)
• –Unlike many efficient vision papers, it does not just trade quality for speed; the reported results stay competitive with top quadratic systems on camera pose, depth, and dense geometry
• –The compact scene state is queryable in real time, which gives the model a practical path from offline reconstruction to interactive and streaming use cases
• –The paper still flags real limits: very long out-of-distribution scenes degrade quality, and queried RGB views remain blurry in high-frequency regions