CrossPressor: Scalable Cross-View Fusion for Feed-Forward 3D Gaussian Splatting

arXiv:2611.06258 · cs.CV · Submitted 2026-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "CrossPressor: Scalable Cross-View Fusion for Feed-Forward 3D Gaussian Splatting".

Jane: Here's the gist of the paper. Feed-forward 3D Gaussian Splatting (3D-GS) becomes increasingly difficult to scale to long multi-view sequences The authors introduce CrossPressor, a training-free method for scalable anchor-based fusion that preserves all decoder-facing anchor queries while constraining the view-growing support sequence through coverage-cycled evidence routing.

Tom: The Representation–Fusion Mismatch The paper identifies a "representation–fusion mismatch," which is the contradiction where bounding the decoder-facing anchor representation does not bound the support evidence required to construct it As more views are assigned to a fixed set of anchors, the representation remains compact while support-side K/V projection and cross-attention continue to grow, exposing a growing support-to-anchor fusion bottleneck even though the decoder-facing representation remains fixed.

Jane: CrossPressor Mechanism CrossPressor addresses this mismatch by routing the view-growing support features into a compact sequence through coverage-cycled evidence routing, which assigns different spatial sampling offsets across support views and varies them across fusion layers.

Tom: This process involves routing before K/V projection and jointly packing routed contexts across fusion layers to reduce the support-side fusion cost without retraining.

Paper summary: Tom: So, looking at CrossPressor: Scalable Cross-View Fusion for Feed-Forward three dee Gaussian Splatting, the authors are proposing this training-free coverage-cycled evidence routing to control the cost and coverage of the multi-view evidence used to construct that representation <ref:2611.06258#pg1,CrossPressor: Scalable Cross-View Fusion for Feed-Forward 3D Gaussian Splatting>.

Jane: It’s really about finding a way to keep things efficient when dealing with long sequences of views by managing that fusion bottleneck they identified.

Lu: The implication is that for scaling feed-forward reconstruction, we need to control both the final representation size and the cost and coverage of the evidence used to build it.

Meng: For us in deployment, if we can do this without retraining, that's a huge win because fine tuning large models is expensive. We need mechanisms that work out of the box on inference time.

Lalam: I’m fairly confident in their direction; they show bounded scaling behavior with the mean routed Key and Value length staying between two thousand four hundred thirty-two and three thousand five hundred forty one tokens across context views <ref:2611.06258#pg2>.

Tom: It seems like they've found a way to keep the representation compact while still getting enough evidence from all those input views without the fusion cost exploding.

Jane: They also showed that this quality-efficiency trade-off holds up even when you apply it to other models, like DepthSplat–ZPressor, maintaining a p5 delta PSNR of negative zero point zero zero seven nine decibels on the ACID dataset.

Lu: That transfer capability is interesting because it suggests this routing idea isn't just a fluke for one specific model architecture. It’s a general principle for managing evidence distribution during fusion.

Meng: If we can deploy something that scales well without needing task-specific training, that makes the whole pipeline much more viable at scale.

Tom: So, CrossPressor gives us a concrete way to manage the computational cost of multi-view evidence in feed-forward three dee Gaussian Splatting <ref:2611.06258#pg1,feed-forward 3D Gaussian Splatting>.

Conclusion: Tom: So, CrossPressor is this training-free method for anchor-based fusion in feedforward three dimensional Gaussian Splatting.

Jane: It's about trying to scale those models to really long sequences of views without blowing up the computation during the fusion part.

Lu: The core idea they are proposing is a way to handle that representation fusion mismatch, where bounding the final representation doesn't automatically control how much support evidence you need.

Meng: They found that as you add more views to fixed anchors, the representation stays small, but the cross attention and K value projection for the support side keep growing huge.

Lalam: CrossPressor fixes that by using this coverage-cycled evidence routing to route those view features into a compact sequence.

Tom: So how does it actually work? What’s the mechanism they used to get that compact sequence?

Jane: They use this coverage-cycled evidence routing, which assigns different spatial sampling offsets across support views and varies those offsets across different fusion layers.

Lu: This routing happens before the K value projection, and they jointly pack all these routed contexts together across the fusion layers. It's a way to reduce that support side fusion cost without needing any retraining at all.

Meng: That sounds good for deployment because no retraining means we don't have to spend months tuning it for every new dataset we throw at it.

Tom: But how much faster is this actually making things run? What are the efficiency gains they’re seeing?

Jane: They claim up to twenty-six point twelve times faster cross view fusion and about one point four three times end to end acceleration on long input sequences.

Lalam: The complexity analysis shows they get rid of the support view count dependence in the main layer specific K value projection and cross attention terms, bounding them by something like big O beta K value times C squared plus T squared C.

Tom: That's a big win for scaling because it removes that dependency on how many views you feed into the fusion block.

Jane: And even though it’s fast, they say the reconstruction quality is still pretty close to what you get with dense fusion. That’s important for visual fidelity.

Lu: They also tested this method on a different model, DepthSplat–ZPressor without retraining and kept the quality degradation very small across different image distributions.

Meng: So it seems like this routing strategy isn't just a fluke for one specific architecture; it’s a general principle for managing evidence distribution in these models.

Tom: The main point is that they found you need to control both the final size of your representation and the cost of the multi-view evidence used to build it.

Jane: CrossPressor suggests that if you want scalable feedforward reconstruction, you have to manage both things simultaneously.

Lalam: It’s a training-free way to control the cost and coverage of that multi-view evidence used for construction.

Tom: It really boils down to using this coverage cycled routing to manage the cost and how much information those views contribute during fusion.

Jane: And it opens up new ways we can think about scaling these complex three dee reconstruction models in the future.

Yuxi Hu, Wenhao Dong, Jun Zhang, Friedrich Fraundorfer

Graz University of Technology · Ghent University

cs.CV

Submitted: 2026-11-04

Updated: 2026-11-04

Comments: Accepted to WACV 2027

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Feed-forward 3D Gaussian Splatting (3D-GS) becomes increasingly difficult to scale to long multi-view sequences The authors introduce CrossPressor…

Key concepts

Representation–Fusion Mismatch
This contradiction occurs when bounding the final anchor representation doesn't limit the necessary support evidence. Even with fixed anchors, fusing growing support features across many views creates a bottleneck because the fusion cost scales with view count, even if the resulting representation size stays small.
Coverage-cycled Evidence Routing
This is a training-free routing mechanism that controls how view-growing features are processed. It assigns different spatial sampling offsets to support views and varies these offsets across fusion layers. This ensures the necessary evidence is routed efficiently without requiring model retraining.
Support-side Fusion Bottleneck
This bottleneck arises because, as more views are added to a fixed set of anchors, the cost associated with projecting and cross-attending support features grows significantly. CrossPressor routes these features into a compact sequence, effectively reducing this fusion cost without changing the anchor representation.
Scaling Behavior
The method demonstrates bounded scaling behavior as context length increases. While dense fusion becomes more expensive, the routed latency remains relatively flat because the routing rule ensures that the mean length of routed K/V tokens stays within a specific, controlled range across different views.

Terminology

Summary

Feed-forward 3D Gaussian Splatting (3D-GS) becomes increasingly difficult to scale to long multi-view sequences The authors introduce CrossPressor, a training-free method for scalable anchor-based fusion that preserves all decoder-facing anchor queries while constraining the view-growing support sequence through coverage-cycled evidence routing. p. 0

The Representation–Fusion Mismatch

The paper identifies a representation–fusion mismatch, which is the contradiction where bounding the decoder-facing anchor representation does not bound the support evidence required to construct it As more views are assigned to a fixed set of anchors, the representation remains compact while support-side K/V projection and cross-attention continue to grow, exposing a growing support-to-anchor fusion bottleneck even though the decoder-facing representation remains fixed.

CrossPressor Mechanism

CrossPressor addresses this mismatch by routing the view-growing support features into a compact sequence through coverage-cycled evidence routing, which assigns different spatial sampling offsets across support views and varies them across fusion layers. This process involves routing before K/V projection and jointly packing routed contexts across fusion layers to reduce the support-side fusion cost without retraining. p. 0

Efficiency Gains

The method achieves up to 26.12× faster cross-view fusion and 1. p. 043× end-to-end acceleration on long input sequences while maintaining reconstruction quality close to dense fusion. p. 0 The complexity analysis shows that CrossPressor removes support-view-count dependence from the dominant layer-specific K/V projection and cross-attention terms, bounding them by O(βKV (TC squared + T 2C)). p. 5

Routing Ablations and Robustness

Experiments show that View+Layer routing provides the best robustness, with a worst-case degradation of -0.0478 dB. The coverage statistics confirm that View-only increases phase coverage across support views while leaving layerwise coverage unchanged, and view-wise staggering is the primary source of robustness. p. 6

Scaling Behavior

The advantage of CrossPressor grows with context length, as dense cross-view fusion becomes progressively more expensive while the routed latency remains nearly flat. The routing rule ensures that the mean routed K/V length remains bounded between 2432. p. 40 and 3541.3 tokens across context views, demonstrating bounded scaling behavior. p. 7

Transfer Capabilities

CrossPressor is applied to DepthSplat–ZPressor without retraining, maintaining a p5 ∆PSNR of-0.0079 dB and a worst-case degradation of-0. p. 60410 dB on the ACID dataset. p. 5 This demonstrates that the quality–efficiency trade-off observed on RE10K persists across different image distributions. p. 7

Conclusion

CrossPressor proposes a training-free coverage-cycled evidence routing to control the cost and coverage of multi-view evidence used to construct the representation, suggesting that scalable feed-forward reconstruction requires controlling not only the final representation size but also the cost and coverage of the multi-view evidence used to construct it.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can scale feed-forward 3D Gaussian Splatting to long multi-view sequences by controlling support-side fusion cost through coverage cycled evidence routing. This allows for up to 26.12 times faster cross-view fusion and 1.43 times end-to-end acceleration on long input sequences while maintaining reconstruction quality close to dense fusion.

  2. The system can maintain a bounded representation size by keeping the decoder facing anchor queries fixed while constraining the view growing support sequence through coverage cycled evidence routing, distributing retained spatial evidence across support views and fusion layers while keeping all support views represented.

  3. The AI system can reduce computational cost without retraining by routing support features before K/V projection, which reduces both support side projection and attention cost, as routing precedes K/V projection, reducing both support-side projection and attention cost without retraining.

Sources

Related papers