CrossPressor sidesteps the fusion bottleneck in long-sequence 3D reconstruction: training-free routing cuts 256-view cross-view fusion from 125 ms to 4.8 ms

When feed-forward 3D Gaussian splatting scales to long multi-view sequences, the bottleneck moves from the decoder-facing representation to support-to-anchor fusion: the anchor count is fixed, but each anchor must aggregate more and more support views. Anchor compression such as ZPressor partitions input views into a fixed number of groups, selects one anchor view per group, and passes only updated anchor features to the Gaussian decoder, so added support views do not directly increase decoder input or Gaussian count. On the 256-view RealEstate10K evaluation, unmodified dense ZPressor takes 125.28 ms for cross-view fusion and 432.16 ms end to end, serving as the reference.

CrossPressor compresses the support-side sequence at inference time while keeping every support camera represented, lowering fusion cost without retraining. It preserves all decoder-facing anchor queries and routes only the view-growing support sequence, so camera grouping, anchor representation, attention modules and the Gaussian decoder stay unchanged. On the same MVSplat–ZPressor checkpoint at 256 views, CrossPressor retains 1.85% of dense QK interactions, reaches 4.80 ms cross-view latency (26.12x speedup) and 303.18 ms end to end (1.43x), with PSNR 23.1203 against 23.1204 for the dense reference.

The routing rule picks a spatial stride per cluster and cycles sampling offsets across support views and fusion layers, so retained evidence covers different spatial locations. Applying the same sparse grid to every support view and fusion layer repeatedly selects the same locations, while selecting only a few complete support views reduces camera participation. In the 256-view routing ablation, view-wise staggering alone improves p5 PSNR degradation from 0.0204 dB under fixed routing to 0.0064 dB, and adding layer cycling improves the worst case to 0.0478 dB while eliminating scenes below 0.05 dB.

Evaluation applies the method at inference to the released MVSplat–ZPressor and DepthSplat–ZPressor checkpoints, with MVSplat–ZPressor as the primary model, using 256 context views and 8 target views on RealEstate10K and a cross-dataset check on 599 ACID scenes; all compared methods share the same checkpoint and no retraining.

The pipeline is not entirely independent of view count: routing metadata still scales with support views, and per-view operations such as feature extraction are unchanged, so end-to-end speedup is smaller than fusion-level speedup. Routed workload is not strictly monotonic because strides are chosen discretely from powers of two and support views are unevenly distributed across anchor clusters; at 128 views four of six clusters still use a finer stride.