Main discovery: MVLoMa is a sparse star-graph matcher that extends LoMa from image pairs to multiple views. It avoids repeating reference-side work across pairwise matches and can exploit complementary geometric evidence from additional observations. (F1)

Method: MVLoMa first uses a frozen multi-view geometry model to jointly describe all images in a star. It injects that cross-view geometric context into local descriptors. (F2)

Matching: The model matches the centre image to all neighbours in a single forward pass. It uses per-view self-attention and bidirectional centre–neighbour cross-attention. (F3)

Supervision: Positive correspondences are derived only from depth-valid, mutually visible dense warps. Training explicitly penalizes high-confidence matches that violate epipolar geometry. (F4)

Concrete results: On 25-view star graphs drawn from HyperSim, ScanNet++, Mip-NeRF 360, and Tanks and Temples, MVLoMa obtained the best macro-average relative-pose AUC among compared methods at the three reported thresholds: 58.65 / 85.41 / 90.99. (F5)

Efficiency: The matcher runs 2.20 times faster than evaluating the 24 centre–neighbour edges sequentially. (F6)

Ablations and encoders: The same MVLoMa design supports four tested frozen geometry encoders—DA3, π^3, VGGT, and VGGT-Ω—with DA3-Giant performing best on the reported encoder-ablation metrics. Controlled ablations isolate the contributions of geometry-aware descriptors, visibility filtering, geometric negative supervision, and joint multi-view inference. (F7, F8)