PocketSplat: Mobile Gaussian Reconstruction via World-Space Latent Allocatio

Wenzhi Guo; Xianda Chen; Dongxuan Chen; Guangchi Fang; Bing Wang

cs.CVOct 2, 2026

PocketSplat: Mobile Gaussian Reconstruction via World-Space Latent Allocatio

Authors: Wenzhi Guo, Xianda Chen, Dongxuan Chen, Guangchi Fang, Bing Wang

Abstract

Mobile Gaussian reconstruction must satisfy two requirements: the reconstruction model must execute within a device resource envelope, and the resulting Gaussian asset must expose a representation size suited to downstream mobile use. Existing feed-forward Gaussian reconstructors commonly decode dense, image-aligned candidates whose final cardinality is implicitly determined by the input resolution and number of views. We present PocketSplat, a feed-forward framework for budgeted mobile Gaussian asset construction. Given a prescribed output budget, PocketSplat organizes dense geometry-aware latent candidates in predicted world space, allocates exact integer capacity across local latent cells, and decodes complete Gaussian attributes only for retained candidates. Cell-conditioned latent fusion aggregates repeated multi-view evidence before decoding, while spatial responsibility decoding adapts Gaussian support after local sparsification. Experiments on DL3DV and out-of-distribution benchmarks establish a strong quality--budget trade-off against feed-forward Gaussian reconstruction baselines. On Mip-NeRF 360, PocketSplat executes directly on a target iPhone and constructs compact, higher-quality Gaussian assets substantially faster than a deployable streamed MVSplat variant; native MVSplat and DepthSplat exceed the device memory budget.

Figures & tables

Figure 1: PocketSplat for mobile Gaussian reconstruction. (a) Multi-view images are converted into a budgeted Gaussian representation and rendered on a mobile device. (b) The DL3DV quality–budget curve shows the reconstruction trade-off across representation budgets; bubble area denotes peak GPU memory.

Figure 2: PocketSplat overview. Multi-view images enter a frozen multi-view transformer, yielding features, depth, and cameras. Each dense candidate stores a latent, reliability, and support. World-space anchors form cells with population, support-floor, and view-mean-detail statistics. Given $B$ , exact-budget allocation returns quotas satisfying $\sum_{c}k_{c}=B$ . Weighted hidden fusion pools each cell, after which retained candidates are decoded and adapted by SRD to form compact Gaussians for novel-view rendering.

		DL3DV			Tanks & Temples			MegaDepth
Method	Gaussians	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
pixelSplat	1,376,256	21.231	0.7119	0.3463	15.839	0.3968	0.5665	16.213	0.5036	0.5079
MVSplat	458,752	20.655	0.6646	0.2694	14.783	0.3471	0.4971	14.779	0.4218	0.4671
DepthSplat	458,752	17.817	0.5370	0.3508	13.722	0.3224	0.5581	14.450	0.4079	0.4869
TranSplat	458,752	20.668	0.6856	0.2630	14.692	0.3423	0.4987	14.880	0.4441	0.4525
F 4 Splat (113K)	111,234	18.541	0.5230	0.3927	12.910	0.3181	0.6146	15.359	0.4230	0.5247

Table 1: Reconstruction quality across in-domain and out-of-distribution datasets. Higher PSNR/SSIM and lower LPIPS are preferable.

Figure 3: Qualitative DL3DV comparison with six examples. Columns show context inputs, MVSplat, DepthSplat, F 4 Splat, PocketSplat at 226K, 338K, and Full, and ground truth.

Figure 4: Tanks & Temples qualitative comparison on Meetingroom, Truck, Church, and Caterpillar. Columns show context inputs, MVSplat, DepthSplat, PocketSplat at 226K, 338K, and Full, and ground truth.

Method / setting	Status	Gaussians	Reconstruction	Inference	Peak	PSNR	SSIM	LPIPS
MVSplat	OOM	458,752	–	–	–	–	–	–
DepthSplat	OOM	458,752	–	–	–	–	–	–
MVSplat-Stream	Deployable	458,752	42.00	40.16	1.99	17.875	0.5793	0.3165
PocketSplat Light	Deployable	112,896	10.67	10.00	2.68	20.018	0.6690	0.2952
PocketSplat Balanced	Deployable	225,792	10.20	9.24	2.89	21.790	0.6343	0.2142
PocketSplat Quality	Deployable	338,000	10.95	9.59	2.64	22.132	0.6503	0.1923

Table 2: On-device Mip-NeRF 360 evaluation. Times are seconds, peak footprint is reported in GB, and higher PSNR/SSIM and lower LPIPS are preferable.

Figure 5: Mobile visualizations on Mip-NeRF 360: Room, Bonsai, and Counter (rows), comparing MVSplat-Stream and PocketSplat at Light, Balanced, Quality, and Dense budgets (columns). PocketSplat budgets are 112,896, 225,792, 338,000, and 451,584, as reported in Table 2 .

Budget	Random	World-space + SRD	No SRD	SRD gain [95% CI]
113,000	18.0134	21.0183	17.6617	3.3567 [3.1792, 3.5340]
226,000	19.0984	22.7899	21.9609	0.8290 [0.7672, 0.8925]
338,000	20.9305	23.1221	22.9514	0.1707 [0.1492, 0.1932]
Full	identity	23.0695	22.9190	0.1505 [0.1163, 0.1847]

Table 3: Quality ablations on DL3DV (PSNR, dB).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Direct path	Final status	Mobile configuration
PocketSplat	Deployable	Deployable	Core ML + ONNX Runtime
MVSplat	OOM	OOM	Direct cost-volume path
MVSplat-Stream	–	Deployable	CPU Core ML + streamed cost volume ( $D=8$ )
DepthSplat	OOM	OOM	Streamed cost volume + staged execution

Appendix

Table 4: Mobile deployment outcomes on the target iPhone under the four-view

256\times 448

FP32 protocol.

Boundary / tensor	Shape
Preprocess image / intrinsics / C2W	$[1,4,3,H,W]$ / $[1,4,3,3]$ / $[1,4,4,4]$
Camera token	$[1,4,1024]$
Backbone state / raw feature	$[1,4,577,1024]$ / $[1,4,577,2048]$
Pre-allocator hidden / depth	$[1,4,32,H,W]$ / $[1,4,H,W]$
Fused hidden to post graph	$[1,N,32]$
Post means / covariance / opacity	$[1,N,3]$ / $[1,N,3,3]$ / $[1,N]$

Appendix

Table 5: Fixed mobile interfaces; batch size is one.

V=4

H=252

W=448

N=451{,}584

. All listed shapes include the batch dimension.

Method	$T_{\mathrm{INFER50}}$ (ms) $\downarrow$	Peak (GB) $\downarrow$
pixelSplat	837.7	8.483
MVSplat	322.8	13.885
DepthSplat	306.2	7.519
TranSplat	537.8	3.552
F 4 Splat (113K)	301.8	17.796
F 4 Splat (226K)	306.4	17.796

Appendix

Table 6: GPU efficiency for the deployment path. Times are median

T_{\mathrm{INFER50}}

in milliseconds and peak memory is reported in GB.

Budget	DBA	BBD	DBA	BBD	reduction
	Attribute projection		Construction		Construction
113,000	1.308	0.411	166.089	163.216	1.730%
226,000	1.289	0.710	174.178	172.276	1.092%
338,000	1.308	1.029	180.567	179.378	0.659%
Full	1.29	1.32	96.75	96.65	0.096%

Appendix

Table 7: Execution-order ablation. Values are median milliseconds per allocation call.

Figure 6: MegaDepth qualitative comparison with five examples. Columns show context inputs, MVSplat, DepthSplat, F 4 Splat, PocketSplat at 226K, 338K, and Full, and ground truth.

Explore similar work

ZipSplat: Fewer Gaussians, Better Splats
Jun 3, 2026Alexander Veicht, Sunghwan Hong, Dániel Baráth +13D Gaussian Splatting Multi-View 3D Reconstruction
ETH Zürich · Microsoft
Monte Carlo Energy Aggregation for Mobile 3D Gaussian Splatting
Jun 29, 2026Xiaobiao Du, YuAn Wang, Hao Li +33DGS Compression 3D Gaussian Splatting
University of Technology Sydney, Australia · Baidu Inc., China · Australian Institute for Machine Learning, Adelaide University, Australia
GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
Apr 16, 2026Roni Itkin, Noam Issachar, Yehonatan Keypur +33D Gaussian Splatting 3DGS Acceleration
The Hebrew University of Jerusalem · Westlake University

Jun 3, 2026J/K move · Enter open · S save

ZipSplat: Fewer Gaussians, Better Splats

Alexander Veicht, Sunghwan Hong, Dániel Baráth, Marc Pollefeys

ETH Zürich · Microsoft

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

Open paper AskPDF arXiv

Feed-forward 3D Gaussian Splatting methods reconstruct a scene from posed or pose-free images in a single forward pass, yet current approaches predict one Gaussian per input pixel, tying the representation budget to camera resolution rather than scene complexity. A flat wall and a richly textured object thus produce equally many Gaussians despite very different geometric needs. We propose ZipSplat, a token-based feed-forward model that decouples Gaussian placement from the pixel grid. A multi-view backbone extracts dense visual tokens, and k-means clustering compresses them into a compact set of scene tokens. Cross- and self-attention refine these tokens, and a lightweight MLP decodes each into a group of Gaussians with unconstrained 3D positions. Because clustering is applied at inference, a single trained model spans the quality-efficiency curve without retraining. ZipSplat operates without ground-truth poses or intrinsics, yet sets a new state of the art on DL3DV and RealEstate10K with

{\sim}6{\times}

fewer Gaussians than pixel-aligned methods, surpassing the best pose-free baseline by 2.1dB and 1.2dB PSNR, respectively. It further generalizes zero-shot to Mip-NeRF360 and ScanNet++, outperforming all comparable baselines. Our project page is at https://veichta.com/zipsplat.

Abstract

Explore similar work

ZipSplat: Fewer Gaussians, Better Splats

Monte Carlo Energy Aggregation for Mobile 3D Gaussian Splatting

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens