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

Yuxi Hu; Wenhao Dong; Jun Zhang; Friedrich Fraundorfer

cs.CVNov 4, 2026

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

Authors: Yuxi Hu, Wenhao Dong, Jun Zhang, Friedrich Fraundorfer

Organizations: Graz University of Technology · Ghent University

Abstract

Feed-forward 3D Gaussian Splatting (3D-GS) becomes increasingly difficult to scale to long multi-view sequences. Anchor-based compression methods such as ZPressor alleviate representation growth by decoding Gaussians only from a compact set of anchor feature maps, but keeping these anchors fixed forces each one to aggregate an increasing number of support views. We identify this representation--fusion mismatch and introduce CrossPressor, a training-free method for scalable anchor-based fusion. CrossPressor preserves all decoder-facing anchor queries 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. Routing before K/V projection, together with joint context packing, further reduces support-side fusion cost without retraining. CrossPressor maintains reconstruction quality close to dense fusion while achieving up to $26.12\times$ faster cross-view fusion and $1.43\times$ end-to-end acceleration on long input sequences. Our project page is at: https://yuhsihu.github.io/CrossPressor.

Figures & tables

Figure 1: Representation–fusion mismatch and CrossPressor. ZPressor [ 25 ] bounds the decoder-facing representation by decoding Gaussians only from a compact set of anchor feature maps. As more context views are assigned to these anchors, the dense support features used to construct each anchor continue to grow. CrossPressor keeps the anchor representation unchanged while routing a bounded support sequence across cameras and spatial locations, with spatial sampling offsets varied across fusion layers. Gray tokens denote dense support evidence, while green tokens indicate routed evidence from different support views.

Figure 2: Overview of CrossPressor. We keep the original ZPressor [ 25 ] pipeline unchanged, including anchor selection, anchor–support grouping, the anchor-query pathway, and the downstream prediction modules. One representative anchor–support group is expanded to compare the fusion stage. The original dense fusion (top) uses all support features as K/V evidence for cross-attention, whereas CrossPressor (bottom) preserves the full anchor-query pathway and routes only the support-side features before K/V projection. For cluster stride $s_{c}$ , the $s_{c}\times s_{c}$ grid enumerates the $s_{c}^{2}$ possible spatial sampling offsets. We index the offset assigned to support view $v$ at fusion layer $\ell$ by $p_{c,v}^{\ell}=(r_{c}(v)+\kappa\ell)\bmod s_{c}^{2}$ , where $r_{c}(v)$ is the support-view rank and $\kappa$ is the layer-wise increment. Moving across support views advances the sampling pattern by one offset, while moving across fusion layers advances it by $\kappa$ , distributing the retained spatial evidence across both dimensions. The selected support features are gathered and packed into a compact sequence before the original K/V projection and cross-attention, producing the same updated-anchor interface for the unchanged downstream prediction pipeline.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	p5 $\Delta$ PSNR $\uparrow$	QK Ret. $\downarrow$	Cross-view ms / speedup	E2E ms / speedup
	Rendering Quality				Attention Load	Efficiency
ZPressor [ 25 ]	23.1204	0.81310	0.18285	–	100.00%	125.28 / 1.00 $\times$	432.16 / 1.00 $\times$
+ ToMe [ 2 ]	23.1205	0.81310	0.18284	-0.0032	1.56%	317.50 / 0.40 $\times$	630.32 / 0.69 $\times$
+ Zero-TPrune [ 22 ]	23.0966	0.81282	0.18306	-0.1666	1.60%	6.91 / 18.12 $\times$	303.14 / 1.43 $\times$
+ CrossPressor	23.1203	0.81311	0.18284	-0.0063	1.85%	4.80 / 26.12 $\times$	303.18 / 1.43 $\times$

Table 1: Training-free cross-view reduction at 256 context views on MVSplat–ZPressor. All methods use the same pretrained checkpoint and preserve the decoder-facing anchor representation and

393{,}216

output Gaussians. Cross-view and E2E entries report latency and speedup relative to Exact. Speedups are computed from the underlying unrounded latency measurements.

Figure 3: Runtime scaling on MVSplat–ZPressor. As the number of context views increases, dense cross-view fusion becomes progressively more expensive, whereas CrossPressor remains nearly flat. The resulting end-to-end speedup reaches

1.43\times

at 256 views.

Figure 4: Fusion-memory scaling on MVSplat–ZPressor and DepthSplat–ZPressor. As the number of context views increases, dense fusion memory grows steadily, whereas the fusion-only incremental peak of CrossPressor remains nearly flat. On MVSplat–ZPressor, the resulting reduction reaches

73.8\%

at 256 views.

Views	Cross-view (ms)	Cross-view speedup $\uparrow$	Fusion mem. (MiB) $\downarrow$
36	17.86 / 4.64	$3.85\times$	120.0 / 162.8
64	31.98 / 4.90	$6.53\times$	163.5 / 170.3
128	62.95 / 5.33	$11.80\times$	302.0 / 182.3
192	93.95 / 4.60	$20.43\times$	484.0 / 164.0
256	125.28 / 4.80	$26.12\times$	653.0 / 170.8

Table 2: Long-context scaling on MVSplat–ZPressor. We report cross-view latency, its speedup over Exact, and the fusion-only incremental peak memory (Fusion mem.) as the number of context views increases. Cross-view latency and fusion memory are reported as Exact / CrossPressor. Speedups are computed from the underlying unrounded latency measurements.

Routing	p5 $\Delta$ PSNR $\uparrow$	Worst $\Delta$ PSNR $\uparrow$	$N_{<-.05}$ $\downarrow$
Fixed	-0.0204	-0.1055	10
View-only	-0.0064	-0.0640	1
Layer-only	-0.0154	-0.1071	2
View+Layer	-0.0063	-0.0478	0

Table 3: Routing decomposition at 256 context views. All variants use identical support views, K/V token counts, and QK interactions, differing only in the sampling phase.

N_{<-.05}

counts scenes with

\Delta

PSNR below

-0.05

dB.

$\beta_{KV}$	QK Ret. $\downarrow$	p5 $\Delta$ PSNR $\uparrow$
0.5	0.57%	-0.0137
1.0	1.46%	-0.0072
1.5	1.86%	-0.0059
2.0	2.28%	-0.0049
3.0	3.85%	-0.0025

Table 4: Sensitivity to the K/V target multiplier at 256 context views. Results are evaluated on a fixed 200-scene subset. QK Ret. denotes the ratio of total retained to dense query–key interactions over this subset, and p5

\Delta

PSNR is the fifth percentile of per-scene PSNR differences relative to Exact.

Views	p5 $\Delta$ PSNR $\uparrow$	Cross-view (ms)	Cross-view speedup $\uparrow$
36	-0.0052	5.76 / 5.44	$1.06\times$
64	-0.0060	6.49 / 5.52	$1.18\times$
128	-0.0110	8.77 / 5.58	$1.57\times$
192	-0.0087	12.67 / 5.64	$2.25\times$
256	-0.0065	16.40 / 5.55	$2.95\times$

Table 5: Transfer to DepthSplat–ZPressor. We apply CrossPressor without retraining. Quality changes are relative to Exact over the same

1{,}034

scenes; cross-view latency is reported as Exact / CrossPressor in milliseconds.

	Rendering Quality			Paired PSNR Robustness
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Mean $\Delta$	p5 $\Delta$	Worst $\Delta$	Failures (.01/.03/.05)
ZPressor [ 25 ]	23.12037	0.813096	0.182849	–	–	–	–
+ ToMe [ 2 ]	23.12052	0.813099	0.182839	+0.0002	-0.0032	-0.1041	6 / 1 / 1
+ Zero-TPrune [ 22 ]	23.09657	0.812823	0.183055	-0.0238	-0.1666	-1.5052	536 / 418 / 334
+ CrossPressor	23.12029	0.813105	0.182836	-0.0001	-0.0063	-0.0478	21 / 2 / 0

Table 6: Detailed rendering-quality comparison at 256 context views on RE10K [ 33 ] . All methods are evaluated on the same

1{,}034

RE10K scenes and

8{,}272

paired target views. Paired PSNR statistics are computed scene-wise relative to Exact. Failures report the number of scenes with

\Delta\mathrm{PSNR}<-0.01/-0.03/-0.05

dB, respectively.

	Latency (ms)				Speedup over Exact
Method	Cross-view $\downarrow$	ZPressor $\downarrow$	Encoder $\downarrow$	E2E $\downarrow$	Cross-view $\uparrow$	ZPressor $\uparrow$	Encoder $\uparrow$	E2E $\uparrow$
ZPressor [ 25 ]	125.28	142.80	418.59	432.16	1.00 $\times$	1.00 $\times$	1.00 $\times$	1.00 $\times$
+ ToMe [ 2 ]	317.50	342.19	616.68	630.32	0.40 $\times$	0.42 $\times$	0.68 $\times$	0.69 $\times$
+ Zero-TPrune [ 22 ]	6.91	14.21	289.49	303.14	18.12 $\times$	10.05 $\times$	1.45 $\times$	1.43 $\times$
+ CrossPressor	4.80	14.12	289.64	303.18	26.12 $\times$	10.11 $\times$	1.45 $\times$	1.43 $\times$

Table 7: Detailed runtime breakdown at 256 context views on RE10K [ 33 ] . We report latency at four nested execution scopes: cross-view fusion, the complete ZPressor module, the encoder, and the end-to-end pipeline. Cross-view entries reproduce the fusion-region measurements used in the main comparison. Speedups are computed from the underlying unrounded latency measurements.

	Rendering Quality			Paired PSNR Robustness
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Mean $\Delta$	p5 $\Delta$	Worst $\Delta$	Failures (.01/.03/.05)
ZPressor [ 25 ]	24.63131	0.763763	0.213089	–	–	–	–
+ ToMe [ 2 ]	24.63146	0.763766	0.213085	+0.0002	-0.0039	-0.0204	3 / 0 / 0
+ Zero-TPrune [ 22 ]	24.60262	0.763369	0.213654	-0.0287	-0.2983	-0.7663	309 / 253 / 205
+ CrossPressor	24.63128	0.763764	0.213084	0.0000	-0.0079	-0.0410	17 / 3 / 0

Table 8: Detailed rendering-quality comparison at 256 context views on ACID [ 16 ] . All methods are evaluated on the same

599

ACID scenes and

4{,}792

paired target views. Paired PSNR statistics are computed scene-wise relative to Exact. Failures report the number of scenes with

\Delta\mathrm{PSNR}<-0.01/-0.03/-0.05

dB, respectively.

	Latency (ms)				Speedup over Exact
Method	Cross-view $\downarrow$	ZPressor $\downarrow$	Encoder $\downarrow$	E2E $\downarrow$	Cross-view $\uparrow$	ZPressor $\uparrow$	Encoder $\uparrow$	E2E $\uparrow$
ZPressor [ 25 ]	125.15	142.77	417.66	427.78	1.00 $\times$	1.00 $\times$	1.00 $\times$	1.00 $\times$
+ ToMe [ 2 ]	319.14	343.74	617.77	627.92	0.39 $\times$	0.42 $\times$	0.68 $\times$	0.68 $\times$
+ Zero-TPrune [ 22 ]	6.96	14.27	289.59	299.64	17.98 $\times$	10.01 $\times$	1.44 $\times$	1.43 $\times$
+ CrossPressor	5.02	14.30	289.11	299.18	24.91 $\times$	9.98 $\times$	1.44 $\times$	1.43 $\times$

Table 9: Detailed runtime breakdown at 256 context views on ACID [ 16 ] . We report latency at four nested execution scopes: cross-view fusion, the complete ZPressor module, the encoder, and the end-to-end pipeline. Speedups are computed from the underlying unrounded latency measurements relative to the same dense ZPressor execution.

Backbone	Views	Exact	Ours	Speedup	Exact	Ours	Speedup	Exact	Ours	Speedup	Exact	Ours	Speedup
		Cross-view			ZPressor			Encoder			End-to-end (E2E)
MVSplat	36	17.86	4.64	$3.85\times$	27.32	13.29	$2.05\times$	178.09	164.24	$1.08\times$	191.97	178.12	$1.08\times$
	64	31.98	4.90	$6.53\times$	42.33	13.70	$3.09\times$	210.05	181.39	$1.16\times$	224.01	195.29	$1.15\times$
	128	62.95	5.33	$11.80\times$	75.60	14.29	$5.29\times$	279.21	218.33	$1.28\times$	291.33	230.59	$1.26\times$
	192	93.95	4.60	$20.43\times$	108.98	13.65	$7.98\times$	348.56	252.96	$1.38\times$	361.05	265.51	$1.36\times$
	256	125.28	4.80	$26.12\times$	142.80	14.12	$10.11\times$	418.59	289.64	$1.45\times$	432.16	303.18	$1.43\times$

Table 10: Detailed runtime scaling on MVSplat–ZPressor and DepthSplat–ZPressor. We report latency for cross-view fusion, the complete ZPressor module, encoder, and end-to-end reconstruction at increasing context lengths. For each backbone, “Exact” denotes the unmodified dense execution of the corresponding released ZPressor model. All latency measurements are in milliseconds. Speedups are computed from the underlying unrounded latency measurements.

Backbone	Views	Exact	Ours	Reduction	Exact	Ours
		Fusion incremental peak			Full-model peak allocated
MVSplat	36	120.0	162.8	$-35.6\%$	3422.8	3424.2
	64	163.5	170.3	$-4.1\%$	3442.8	3444.1
	128	302.0	182.3	$39.7\%$	3490.9	3492.0
	192	484.0	164.0	$66.1\%$	4194.5	4195.9
	256	653.0	170.8	$73.8\%$	5570.5	5572.1

Table 11: Detailed memory scaling on MVSplat–ZPressor and DepthSplat–ZPressor. We report the fusion-only incremental peak memory used in the scaling analysis, together with the full-model peak allocated memory. For each backbone, “Exact” denotes the unmodified dense execution of the corresponding released ZPressor model. All values are reported in MiB. Reduction is computed as

1-\mathrm{CrossPressor}/\mathrm{Exact}

for the fusion-only peak from the underlying unrounded memory measurements; negative values therefore indicate additional routing and packing workspace at short contexts.

Views	Mean support views/cluster	Mean routed K/V length	Stride count			QK Ret. $\downarrow$	Cross-view (ms) Exact / CrossPressor
			$s_{c}=2$	$s_{c}=4$	$s_{c}=8$
36	5.00	2432.0	3	3	0	11.88%	17.86 / 4.64
64	9.67	2858.7	1	5	0	7.22%	31.98 / 4.90
128	20.33	3541.3	0	4	2	4.25%	62.95 / 5.33
192	31.00	2464.0	0	1	5	1.94%	93.95 / 4.60
256	41.67	2890.7	0	1	5	1.69%	125.28 / 4.80

Table 12: Routing-load scaling for the MVSplat–ZPressor runtime benchmark. Routing statistics are exact offline counts for the corresponding runtime-scaling workloads.

\widetilde{N}_{c}

denotes the routed support-token length per cluster. Stride columns report the number of the six anchor clusters assigned to each stride. Cross-view latency is reported as Exact / CrossPressor in ms.

Routing	View Cov.	Layer Cov.	Mean $\Delta$ PSNR $\uparrow$	p5 $\Delta$ PSNR $\uparrow$	Worst $\Delta$ PSNR $\uparrow$	$N_{<-.01}$ $\downarrow$	$N_{<-.03}$ $\downarrow$	$N_{<-.05}$ $\downarrow$
Fixed	2.41%	1.85%	+0.0027	-0.0204	-0.1055	140	24	10
View-only	74.11%	1.85%	+0.0005	-0.0064	-0.0640	25	2	1
Layer-only	2.41%	11.10%	+0.0005	-0.0154	-0.1071	100	15	2
View+Layer	74.11%	11.10%	-0.0001	-0.0063	-0.0478	21	2	0

Table 13: Full routing decomposition at 256 context views. All variants use identical support views, K/V token counts, and QK interactions, differing only in spatial phase assignment (

\kappa=5

for layer cycling). View Cov. measures the fraction of spatial phases covered across support views within a cluster and layer, averaged equally over cluster–layer pairs; Layer Cov. measures the corresponding coverage of each support view across the six fusion layers. Quality statistics are computed from per-scene

\Delta

PSNR relative to Exact over

1{,}034

scenes, with eight target views averaged within each scene.

N_{<-\tau}

counts scenes with

\Delta

PSNR below

-\tau

dB.

$\kappa$	Mean $\Delta$ PSNR $\uparrow$	p5 $\Delta$ PSNR $\uparrow$	Worst $\Delta$ PSNR $\uparrow$	$N_{<-.01}$ $\downarrow$
1	+0.0001	-0.0045	-0.0333	1
3	-0.0002	-0.0066	-0.0239	2
5	-0.0002	-0.0059	-0.0143	2
7	-0.0002	-0.0055	-0.0090	0

Table 14: Sensitivity to the layer-wise phase increment at 256 context views. Results are evaluated on the same fixed 200-scene subset used for the K/V sensitivity study. All settings use identical routed token counts and phase-coverage statistics; only

\kappa

is varied.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,043 papers

Oct 15, 2026

Anchor3DGS: Feed-Forward 3D Gaussian Splatting with Compact Anchor-Based Representation

Recent feed-forward 3D Gaussian Splatting (3DGS) methods have enabled effective multi-view 3D reconstruction by predicting Gaussian primitives in a single forward pass. However, existing approaches predominantly adopt a pixel-aligned paradigm that predicts one Gaussian per input pixel, coupling the Gaussian count to the input resolution and number of views. This also leads to redundant primitives in textureless regions and insufficient coverage in geometrically complex areas. We present Anchor3DGS, a pose-free feed-forward 3DGS framework that replaces pixel-aligned prediction with a compact anchor-based representation. Guided by information entropy, our method places a set of 3D anchors in the scene, concentrating them in regions of higher visual complexity while maintaining broad spatial coverage. Each anchor performs occlusion-aware aggregation of multi-view image features, exchanges information with other anchors, and is decoded into a set of local Gaussians. A lightweight feature enhancement module further refines the coarse RGB output from Gaussian splatting and improves rendering fidelity. Extensive experiments on diverse benchmark datasets demonstrate that our Anchor3DGS achieves state-of-the-art novel view synthesis quality while only using an order of magnitude fewer Gaussian primitives.

Sheng Ye, Jiangke Lin, Fudong Wang +3

Tsinghua University · Ant Group, Inclusion AI

3D Gaussian Splatting Multi-View 3D Reconstruction Novel View Synthesis Feed-Forward 3D Reconstruction

Ask PDF

Sep 27, 2026

FeCoSplat: Feedback-Guided Compression for Feed-Forward 3D Gaussian Splatting

Feed-forward 3D Gaussian Splatting (3DGS) enables efficient novel-view synthesis from sparse multi-view images, yet its representations remain costly to store and transmit. Existing approaches compress either the input images, incurring heavy receiver-side reconstruction, or the reconstructed Gaussian primitives, which are difficult to compress due to their heterogeneous and irregular attributes. We instead compress compact intermediate features, providing a better balance between compression efficiency and receiver-side complexity. Based on this paradigm, we propose FeCoSplat, a feedback-guided compression framework for feed-forward 3DGS. FeCoSplat first compresses multi-view features to obtain an intermediate 3DGS, whose rendered views are used as feedback to guide a second-stage compression for further refinement. The resulting bitstreams are decoded into a compact implicit state, from which the final Gaussian primitives are reconstructed with a lightweight predictor. Experiments demonstrate that FeCoSplat achieves favorable rate--distortion performance, particularly at low bitrates, while requiring only 3.45M parameters for receiver-side Gaussian reconstruction. Code will be released soon.

Yuxuan Li, Yihang Chen, Yufeng Zhang +2

Shanghai Jiao Tong University · Monash University

3DGS Compression 3D Gaussian Splatting Feed-Forward 3D Reconstruction Gaussian Splatting Compression

Ask PDF

Sep 5, 2026

AVSplat: Dense-View Feed-Forward 3D Gaussian Splatting with Assist-View Preconditioning

Pose-free feed-forward 3D Gaussian Splatting enables novel view synthesis from uncalibrated multi-view images. Although more views should improve performance, existing methods often degrade with dense-view inputs because global aggregation spreads attention over many tokens, and naive voxel fusion averages many Gaussians into overly smooth representations. We present AVSplat, a framework that turns additional views into reliable signals for both aggregation and representation. Before global attention, each view performs a single lightweight interaction with a small set of Assist Views chosen for relevance and diversity, and the cached features provide a focused scene context that stabilizes correspondence. For representation, we use adaptive temperature-aware voxel fusion that sharpens attribution under high occupancy, guided by occupancy and point confidence. Crucially, AVSplat restores positive view scaling where performance remains stable or improves as more input views are added, instead of degrading in the dense-view regime. Ablations show that Assist View Preconditioning is primarily responsible for preventing dense-view degradation, while Occupancy-guided Voxel Fusion contributes most of the single-point image-quality gains.

Muyu Xu, Fangneng Zhan, Yu Wei +2

Nanyang Technological University, Singapore 639798, Singapore · Harvard University, MA 02138, USA

3D Gaussian Splatting Novel View Synthesis Feed-Forward 3D Reconstruction

Ask PDF

Abstract

3D Gaussian Splatting

Explore similar work

Anchor3DGS: Feed-Forward 3D Gaussian Splatting with Compact Anchor-Based Representation

FeCoSplat: Feedback-Guided Compression for Feed-Forward 3D Gaussian Splatting

AVSplat: Dense-View Feed-Forward 3D Gaussian Splatting with Assist-View Preconditioning