TileSkipper: Region-Adaptive Tile Pruning for 3D Gaussian Splatting

Jingxing Li; Yongjae Lee; Deliang Fan; Abhay Kumar Yadav; Cheng Peng; Rama Chellappa

cs.CVOct 7, 2026

TileSkipper: Region-Adaptive Tile Pruning for 3D Gaussian Splatting

Authors: Jingxing Li, Yongjae Lee, Deliang Fan, Abhay Kumar Yadav, Cheng Peng, Rama Chellappa

Organizations: School of Electrical, Computer and Energy Engineering Arizona State University Tempe, AZ, USA · Department of Electrical and Computer Engineering Johns Hopkins University Baltimore, MD, USA

Abstract

Tiled 3D Gaussian Splatting rasterizers often use one scene-wide contribution cutoff for tile enumeration, although content differs in its sensitivity to support truncation. TileSkipper selects a static per-Gaussian cutoff policy for a frozen checkpoint. Calibration renders measure candidate pair savings and an isolated-removal distortion proxy that accounts for front transmittance and background color. The method allocates cutoffs across 64 Gaussian groups and accepts policies only after complete renders on disjoint selection views. The exported policy uses one byte per Gaussian, with no parameter updates, additional kernel, or per-frame policy inference. Across 13 scenes from Mip-NeRF 360, Tanks & Temples, and Deep Blending, a fixed-policy AccuTile sweep gives dataset-macro speedups of $1.088\times$ at standard resolution and $1.238\times$ at 3840 pixels wide, with $-0.007/-0.023$ dB mean PSNR change. Six integrations with existing opacity-aware bounds yield $1.009\times$ -- $1.121\times$ compiler-only speedups. For four ports from $3σ$ rasterizers, we separately attribute the prior exact-bound transition and our incremental gain. Matched-quality ablations show modest gains over scene-global calibration and parity with per-Gaussian control; the standalone comparison with AdaGScale is regime-dependent.

Figures & tables

Figure 1: A scene-global cutoff leaves allocation slack. (a) On this diagnostic work–fidelity frontier, the compiled regional policy reaches

16.4

dB higher teacher PSNR than the swept scene-global cutoff at the same pair count; this is the median scene by that margin. The number measures approximation fidelity to the same backbone under equal enumerated work, not a

16.4

dB improvement in ground-truth PSNR. At matched teacher PSNR, Regional is

1.040

–

1.046\times

faster than Ours-Global at standard resolution and

1.041

–

1.083\times

faster at 4K ( Table 2 ); it ties the per-Gaussian variant. (b) Work-weighted quartiles (thick lines and dots) and full selected ranges (thin lines) over

64

regions. Eleven of thirteen scenes select more than one level; seven span the full

\tau_{0}

–

8\tau_{0}

range, and two collapse to one level. This dispersion shows what a scene-global constant cannot represent; it does not by itself measure end-to-end speedup.

Figure 2: TileSkipper changes pair enumeration at stage 2. The blue and orange outlines denote support under the baseline and compiled cutoffs. Tiles excluded by the compiled support generate no Gaussian–tile pair and therefore enter neither sorting nor blending. For tile

t

, the figure shows the resulting depth order and compositing expression;

T_{i}

is transmittance before contribution

i

Figure 3: Offline compilation and runtime application of TileSkipper. The diagram abbreviates the region-count symbol

K_{\mathrm{reg}}

K

. (a) Once per scene, the compiler uses a frozen 3DGS checkpoint and calibration views to measure per-Gaussian pair savings

s

and isolated-removal risk proxies

\rho

for each threshold level. The backward probe collects sensitivities without updating any model parameter. Gaussians are assigned to eight efficiency-quantile strata

q_{1},\ldots,q_{8}

(low to high). Within each stratum, Morton ordering and workload-balanced contiguous partitioning produce

K_{\mathrm{reg}}=64

regions in total; Gaussian counts may differ between regions. For region

c

and level

\ell

S_{c,\ell}=\sum_{g\in c}s_{g,\ell}

and

R_{c,\ell}=\sum_{g\in c}\rho_{g,\ell}

aggregate pair savings and risk. Each penalty

\lambda

selects one level for each of the

K_{\mathrm{reg}}

regions. The 48 resulting proposals and uniform-level policies are rendered on disjoint held-out training views. A policy passes only if its minimum per-view teacher PSNR meets

P_{\min}

and its worst-view p99 absolute residual does not exceed

E_{\max}

; among passing policies, the compiler selects the one with the greatest pair savings. Teacher comparisons use the unmodified backbone; p99 is computed from maximum absolute RGB-channel residuals on a fixed

4\times

spatial subsample. Test views are never used for selection. (b) Each Gaussian inherits its region’s selected level, exported as one uint8 per Gaussian plus a threshold lookup table. At inference, the default threshold is

\tau_{g}=m_{g}\tau_{0}

, where

m_{g}

is the selected multiplier and

\tau_{0}

is the baseline threshold. Backend-specific handling of opacity caps is described in Section E.1 . The contribution field

a_{g}(x)

defines the thresholded support

\{x:a_{g}(x)\geq\tau\}

. A Gaussian–tile pair is created when the tile intersects this support, using

\tau_{0}

for the baseline and

\tau_{g}

for TileSkipper. The dashed blue contour denotes the original support, the orange contour the compiled support, and shaded orange tiles the skipped pairs. Gaussian parameters remain unchanged. The byte and lookup table are read inside tile enumeration, with no per-frame policy inference or additional kernel; projection, sorting, and blending remain unchanged. Footprints, partitions, and table entries are schematic.

Base model/backend	Integration	Scenes	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS $\downarrow$	$R$	+Ours speedup [95% CI] (wins)
Native and native-compatible rasterizer integrations
vanilla 3DGS ( Kerbl et al., 2023 )	3DGS CUDA	13	$-0.003$	$-0.000069$	$-0.000092$	$0.330$	$\mathbf{1.493\times}$ $[1.350,1.656]$ (13/13)
Mini-Splatting ( Fang and Wang, 2024 )	3DGS CUDA	13	$-0.008$	$-0.000104$	$-0.000172$	$0.435$	$\mathbf{1.346\times}$ $[1.260,1.451]$ (13/13)
Speedy-Splat ( Hanson et al., 2025a )	native	13	$-0.007$	$-0.000109$	$-0.000202$	$0.825$	$\mathbf{1.096\times}$ $[1.078,1.116]$ (13/13)
FastGS ( Ren et al., 2025 )	native	13	$-0.004$	$-0.000088$	$-0.000098$	$0.979$	$\mathbf{1.009\times}$ $[1.005,1.014]$ (11/13)
SeeLe ( Zhu et al., 2026 )	native transfer	13	$-0.008$	$-0.000120$	$-0.000073$	$0.800$	$\mathbf{1.091\times}$ $[1.070,1.115]$ (13/13)

Table 1: Base model/backend versus the same path plus Ours (

K_{\mathrm{reg}}=64

). Every row has thirteen scenes and twelve balanced same-process timing cycles. All quality deltas are Ours minus the named base (higher PSNR/SSIM and lower LPIPS are better).

R

is the mean pair ratio; intervals bootstrap scenes and parentheses count wins. SeeLe and FlashGS transfer a Speedy-Splat policy unchanged into the named native rasterizer. LightGaussian and Taming-3DGS vary the checkpoint on one common AccuTile backend; Taming’s native absolute-opacity activation is preserved. The vanilla 3DGS, Mini-Splatting, 2DGS, and Scaffold-GS rows include the prior exact-bound transition, separated in Tables 14 and 15 .

Resolution	Teacher PSNR	Regional / Ours-Global	Per-G / Ours-Global	Regional / per-G	Wins
Standard	60 dB	$\mathbf{1.0397\times}$ $[1.0290,1.0508]$	$1.0435\times$	$0.9963\times$	13/13
	55 dB	$\mathbf{1.0457\times}$ $[1.0373,1.0551]$	$1.0460\times$	$0.9997\times$	13/13
	50 dB	$\mathbf{1.0401\times}$ $[1.0313,1.0504]$	$1.0332\times$	$1.0067\times$	13/13
4K	60 dB	$\mathbf{1.0413\times}$ $[1.0208,1.0612]$	$1.0556\times$	$0.9864\times$	10/13
	55 dB	$\mathbf{1.0709\times}$ $[1.0500,1.0923]$	$1.0811\times$	$0.9906\times$	13/13
	50 dB	$\mathbf{1.0825\times}$ $[1.0632,1.1029]$	$1.0816\times$	$1.0008\times$	13/13

Table 2: Strict matched-quality ablation of compiler granularity on thirteen scenes. Each variant’s own (distortion, latency) Pareto envelope is interpolated at a common teacher-PSNR target against the unmodified backbone. Ours-Global is our compiler with one decision (

K_{\mathrm{reg}}=1

), not an external baseline; Regional is the full

K_{\mathrm{reg}}=64

method. Brackets are

95\%

scene-bootstrap intervals. Regional improves on Ours-Global everywhere and ties the per-Gaussian variant. Thus

K_{\mathrm{reg}}

is a decision regularizer and search granularity, not a demonstrated source of speed or storage reduction.

Dataset	Scenes	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS $\downarrow$	$R$	TileSkipper speedup [95% CI] (wins)
(a) Native resolution
All scenes	13	$+0.0059$	$+0.000021$	$-0.000337$	$0.945$	$1.010\times$ $[0.990,1.025]$ (10/13)
Mip-NeRF 360	9	$+0.0056$	$+0.000066$	$-0.000349$	$0.934$	$1.017\times$ $[1.004,1.030]$ (7/9)
Tanks & Temples	2	$-0.0001$	$-0.000032$	$-0.000161$	$1.008$	$0.960\times$ $[0.920,1.002]$ (1/2)
Deep Blending	2	$+0.0138$	$-0.000126$	$-0.000461$	$0.930$	$1.027\times$ $[1.024,1.029]$ (2/2)
(b) 4K

Table 3: Standalone comparison with AdaGScale at native and 4K resolution. Both methods use the same Speedy-Splat checkpoints and AdaGScale FlashGS renderer, with policies selected under common training-view quality constraints (details in Section C.2 ). Deltas are TileSkipper minus AdaGScale;

R

is the TileSkipper/AdaGScale pair ratio, and speedup is AdaGScale/TileSkipper latency. Brackets are 95% scene-bootstrap intervals; parentheses count speed wins. The two-scene intervals are descriptive.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Family	Stage	Control unit	Parameter updates
Speedy-Splat (AccuTile) / FastGS [ Hanson et al., 2025a , Ren et al., 2025 ]	Tile-bounding	Whole scene, hand-set	—
AdR-Gaussian / FlashGS / QuadBox [ Wang et al., 2024 , Feng et al., 2025 , Li et al., 2026 ]	Tile geometry	Per projected Gaussian	—
AdaGScale [ Jo et al., 2026 ]	Tile-bounding	Per view, per Gaussian	—
SeeLe [ Zhu et al., 2026 ]	View prefetch + blending	Camera cluster	fine-tuning
Fragment Pruning [ Ye et al., 2024 ]	Per-pixel rasterization	Per-Gaussian	learned
Gaussian-set pruning [ Lee et al., 2024 , Girish et al., 2025 , Niemeyer et al., 2025 , Hanson et al., 2025b ]	Gaussian set	Per-Gaussian	varies

Appendix

Table 4: Positioning by rasterization stage and decision unit. TileSkipper compiles heterogeneous, static per-Gaussian cutoffs without checkpoint updates or a single hand-set scene-wide cutoff; complete held-out renders validate each proposed policy.

Figure 4: Measured headroom under two controlled interventions. (a) Stage shares measured with the rasterizer’s own CUDA events. (b) The mechanisms have different primary targets, but their runtime effects are not assumed to be disjoint or additive. Thresholding changes tile enumeration, duplicate/sort work, and correlated blending; pre-culling avoids processing rejected Gaussian blocks and may also change downstream work. Within each profiled scene, the duplicate-plus-sort share grows with resolution:

21.4\%\!\rightarrow\!34.2\%

for bicycle and

31.6\%\!\rightarrow\!41.6\%

for counter . The estimated ceiling of the tested perfect, free visibility culler falls from

22.6\%

7.6\%

and from

9.3\%

2.2\%

, respectively. Dispatch uses measured net benefit, including planning overhead.

Figure 5: Each mechanism has a regime, and the cost model names it. (a) The compiled policy strengthens with resolution; both

P

and

|\Omega|

change in this sweep. The separate stage profiles in Figure 4 show the within-scene changes and do not conflate the bicycle and counter endpoints. (b) Pre-culling strengthens as the measured

aN/T

share grows. The dense bars are stress-test checkpoints; the two legacy packaged checkpoints are not used as evidence for any named model family. On Speedy-Splat’s

0.4

M checkpoint the frustum test costs more than it saves and “both” is worse than the policy alone, which is the case the dispatch rule catches.

Width	MP/frame	$R$	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS $\downarrow$	Speedup (dataset macro / scene geom.)
Standard	$1.15$	$0.8248$	$-0.0074$	$-0.000109$	$-0.000202$	$1.0879\times/1.0955\times$
$1920$ px	$2.38$	$0.8044$	$-0.0141$	$-0.000416$	$-0.000145$	$1.1283\times/1.1334\times$
$2560$ px	$4.23$	$0.7908$	$-0.0166$	$-0.000532$	$-0.000167$	$1.1615\times/1.1665\times$
$3840$ px	$9.51$	$0.7750$	$-0.0231$	$-0.000742$	$-0.000221$	$1.2384\times/1.2495\times$

Appendix

Table 5: Fixed-policy resolution stress test on AccuTile, all thirteen scenes. The same deployable

K_{\mathrm{reg}}=64

policies, selected using held-out training views at standard resolution, are applied unchanged at every width; test views are used only for evaluation. Higher-resolution rows bilinearly resample the same source views and therefore test workload scaling rather than natively captured high-resolution content. Frozen per-Gaussian attributes are cached once in both arms. Timing rotates AccuTile, Ours-Global , and Ours-Regional in one process with three warmup and seven measured repetitions per view. Speedup is AccuTile/Regional latency; “dataset macro” equally weights the three dataset-level arithmetic means, while “scene geom.” is the geometric mean over all thirteen scenes.

R

and all quality deltas are arithmetic scene means over full official test sets; deltas are Regional minus AccuTile (higher PSNR/SSIM and lower LPIPS are better).

$K_{\mathrm{reg}}$	mean $\Delta$ PSNR	$R$	vs. Official-Fixed	vs. $K_{\mathrm{reg}}=1$ [95% CI]	compile (s)
$1$	$-0.0041$	$0.9192$	$1.0484\times$	$1.0000\times$	$2.8$
$2$	$-0.0047$	$0.8868$	$1.0641\times$	$1.0165\times$ $[1.0105,1.0223]$	$3.8$
$4$	$-0.0051$	$0.8479$	$1.0843\times$	$1.0351\times$ $[1.0276,1.0424]$	$4.9$
$8$	$-0.0067$	$0.8329$	$1.0894\times$	$1.0412\times$ $[1.0298,1.0525]$	$7.0$
$16$	$-0.0054$	$0.8306$	$1.0921\times$	$1.0422\times$ $[1.0289,1.0551]$	$7.7$
$32$	$-0.0064$	$0.8273$	$1.0949\times$	$1.0436\times$ $[1.0316,1.0555]$	$8.7$

Appendix

Table 6: Region-count ablation on Speedy-Splat, all thirteen scenes.

K_{\mathrm{reg}}=1

is our compiler restricted to one scene-level decision; larger

K_{\mathrm{reg}}

increases the number of policy groups, with spatial workload balancing within each risk stratum. Every row uses the same ladder, calibration/selection views, and quality constraints; the existing construction uses

S_{\mathrm{eff}}=\min(8,K_{\mathrm{reg}})

risk strata. Latency uses the same honest-cache, rotating three-arm protocol as Table 1 ; brackets are

95\%

scene-bootstrap intervals. We retain the globally fixed

K_{\mathrm{reg}}=64

; the added

K_{\mathrm{reg}}=2,4,512,1024

rows use the same three-arm evaluation protocol.

$K_{\mathrm{reg}}$	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS $\downarrow$	$R$	vs. exact bound	vs. $K_{\mathrm{reg}}=1$ [95% CI]	compile (s)
$1$	$-0.0049$	$-1.21\!\times\!10^{-4}$	$-7.07\!\times\!10^{-5}$	$0.9161$	$1.0838\times$	$1.0000\times$	$11.7$
$2$	$-0.0056$	$-1.34\!\times\!10^{-4}$	$-8.94\!\times\!10^{-5}$	$0.9027$	$1.0974\times$	$1.0144\times$ $[1.0079,1.0218]$	$18.0$
$4$	$-0.0068$	$-1.61\!\times\!10^{-4}$	$-1.00\!\times\!10^{-4}$	$0.8882$	$1.1129\times$	$1.0287\times$ $[1.0169,1.0432]$	$24.8$
$8$	$-0.0075$	$-1.67\!\times\!10^{-4}$	$-1.11\!\times\!10^{-4}$	$0.8801$	$1.1200\times$	$1.0334\times$ $[1.0197,1.0485]$	$37.7$
$64$	$-0.0093$	$-1.83\!\times\!10^{-4}$	$-1.06\!\times\!10^{-4}$	$0.8787$	$1.1216\times$	$1.0349\times$ $[1.0215,1.0496]$	$53.5$
$256$	$-0.0098$	$-1.93\!\times\!10^{-4}$	$-1.09\!\times\!10^{-4}$	$0.8782$	$1.1228\times$	$1.0360\times$ $[1.0214,1.0520]$	$57.9$

Appendix

Table 7: Region-count ablation on 2DGS, all thirteen scenes. All variants use the same frozen 30k-iteration 2DGS checkpoints at native resolution, the same 16 calibration and 32 disjoint selection views, the 11-level ladder,

S_{\mathrm{eff}}=\min(8,K_{\mathrm{reg}})

risk strata, and 48 Lagrangian penalties with the same proposal-generation procedure. Selection requires minimum-view teacher PSNR

\geq 48

dB and worst-view p99

\leq 0.012

; these are common constraints, not exact matching of attained distortion. The baseline is the 2DGS path with the exact

1/255

opacity bound, so the speedups exclude the prior

3\sigma

-to-exact-bound transition. Quality deltas and pair ratio

R

are relative to this baseline on the full test views. Each measurement batch has five arms with cached frozen activations and fifteen balanced same-process timing cycles. The added low-

K_{\mathrm{reg}}

and high-

K_{\mathrm{reg}}

batches each include the same frozen

K_{\mathrm{reg}}=1

and

64

controls and the exact-bound baseline; each ratio uses its own batch’s timing. Speedups are scene geometric means; quality deltas,

R

, and compilation times are arithmetic scene means. Brackets are 95% scene-bootstrap intervals. Compilation time excludes checkpoint loading.

Grouping	Risk	Spatial	Work-bal.	Work CV	$R$	vs. AccuTile	Full/row [95% CI] (wins)
Random	no	no	yes	$0.005$	$0.8832$	$1.0649\times$	$1.0360\times$ $[1.0285,1.0458]$ (13/13)
Uniform grid	no	yes	no	$3.652$	$0.8823$	$1.0658\times$	$1.0351\times$ $[1.0293,1.0408]$ (13/13)
No risk strata	no	yes	yes	$0.005$	$0.8480$	$1.0820\times$	$1.0196\times$ $[1.0151,1.0243]$ (13/13)
No workload balance	yes	yes	no	$0.684$	$0.8087$	$1.1034\times$	$0.9998\times$ $[0.9985,1.0011]$ (6/13)
Full	yes	yes	yes	$0.047$	$0.8076$	$1.1032\times$	–

Appendix

Table 8: Region-construction ablation on all thirteen Speedy-Splat scenes. We fix

K_{\mathrm{reg}}=64

decisions and keep probes, ladder, and search unchanged. Each validation-selected Pareto curve is interpolated in teacher MSE to 55 dB, with only its two bracketing policies timed in a common balanced process. Random retains work balancing alone; Uniform grid retains spatial locality alone; No-risk keeps Morton locality and work balancing; No-workload keeps risk strata and Morton order but partitions by Gaussian count. Work CV measures the coefficient of variation in baseline pair workload across regions. Full/row above one indicates that Full is faster, and brackets give

95\%

scene-bootstrap intervals.

Threshold	mean $\Delta$ PSNR	worst $\Delta$ PSNR	worst scene	vs. Official-Fixed
$1.25\tau_{0}$	$-0.0011$	$-0.0030$	bonsai	$1.0194\times$
$1.5\tau_{0}$	$-0.0038$	$-0.0126$	bonsai	$1.0305\times$
$2\tau_{0}$	$-0.0126$	$-0.0493$	bonsai	$1.0530\times$
$3\tau_{0}$	$-0.0425$	$-0.2005$	bonsai	$1.0855\times$
$4\tau_{0}$	$-0.0853$	$-0.4255$	bonsai	$1.1113\times$
$6\tau_{0}$	$-0.1900$	$-0.9404$	bonsai	$1.1497\times$

Appendix

Table 9: “Why not simply raise the constant?” Each row applies one fixed multiple of

\tau_{0}

to all thirteen scenes in one dedicated cutoff sweep. The mean is reassuring and the worst scene is not: the same scene ( bonsai ) is the least favourable at every multiple, so a single constant must be clamped to it. Matching Ours-Regional ’s worst-scene quality caps a fixed constant at

1.62\tau_{0}

and

1.036\times

; matching its

1.1009\times

speed in this sweep forces the constant to

3.6\tau_{0}

, where the worst scene loses

-0.334

dB against

-0.022

dB. The independent all-view twelve-cycle main retiming is

1.096\times

. Choosing safely requires measuring the scene—the compilation step already present in Ours-Global .

Calibration views	$\Delta$ PSNR	$R$	vs. AccuTile [95% CI] (wins)	$16$ /row [95% CI] (wins)	Compile (s)
$4$	$-0.0095$	$0.8661$	$1.0827\times$ $[1.0579,1.1079]$ (13/13)	$1.0036\times$ $[0.9781,1.0257]$ (9/13)	$12.0$
$8$	$-0.0109$	$0.8342$	$1.0774\times$ $[1.0260,1.1197]$ (11/13)	$1.0085\times$ $[0.9831,1.0358]$ (7/13)	$15.4$
$\mathbf{16}$	$-0.0099$	$0.8076$	$1.0866\times$ $[1.0542,1.1160]$ (11/13)	–	$9.0$
$32$	$-0.0074$	$0.8000$	$1.0889\times$ $[1.0523,1.1238]$ (11/13)	$0.9979\times$ $[0.9792,1.0229]$ (3/13)	$17.2$

Appendix

Table 10: Calibration-view sensitivity on all thirteen Speedy-Splat scenes. Each family uses

K_{\mathrm{reg}}=64

and 32 disjoint selection views; only the number of calibration views changes. Curves are interpolated to 55 dB teacher PSNR and their bracketing policies are timed together. Brackets bootstrap scenes and parentheses count wins.

Scene	Frames	Valid flow	$R$ Global	$R$ Regional	$E$ Global	$E$ Regional	$E_{n}$ Regional/Global
Bicycle	97	45.8%	0.922	0.815	2.700	1.798	0.926 $\times$
Bonsai	145	77.4%	0.898	0.772	2.081	1.373	0.998 $\times$
Counter	117	79.6%	0.803	0.738	1.386	1.549	1.050 $\times$
Aggregate	359	69.6%	0.873	0.775	1.982	1.564	0.990 $\times$

Appendix

Table 11: Dense, flow-aligned temporal evaluation. Before constructing test-camera paths, we select Ours-Global and Ours-Regional near 55 dB on held-out training views. We insert three SE(3)-interpolated views between adjacent test poses, masking occlusions with bidirectional RAFT-Small consistency. The flow-aligned residual-change RMSE is

E

(

\times 10^{-3}

);

E_{n}

normalizes it by each policy’s frame RMSE. Lower

E

and Regional/Global are better. Aggregates are geometric means, with valid flow weighted by transition.

Rasterizer	Resolution	Scenes	$R$	Teacher PSNR	vs. Official-Fixed	vs. Ours-Global
SeeLe	Standard	13	0.8003	54.75	$\mathbf{1.0908\times}$	$1.0437\times$
SeeLe	4K ( $3840$ px)	13	0.7539	56.41	$\mathbf{1.2602\times}$	$1.1233\times$
FlashGS	Standard	13	0.8205	52.52	$\mathbf{1.1214\times}$	$1.0605\times$
FlashGS	4K ( $3840$ px)	13	0.7723	50.38	$\mathbf{1.3166\times}$	$1.1379\times$

Appendix

Table 12: Ours-Global and Ours-Regional policies compiled on AccuTile and transferred unchanged into two official optimized rasterizers. FlashGS retains its precise intersection and scheduling optimizations.

R

and teacher PSNR are measured against each rasterizer’s native path. Every row uses twelve balanced same-process timing cycles; no row recalibrates or fine-tunes either policy.

Figure 6: A prediction made before the implementation, and then tested. (a) Fitting only

a

— from a single oracle-subset comparison taking seconds — predicts the measured ceiling of region pre-culling to a mean absolute error of

0.0040

over thirteen scenes. (b) The prediction is unfavourable: the per-Gaussian term is

14.3\%

of the honest frame on average, so the tested perfect, free oracle is capped near

1.17\times

. The estimate is specific to this checkpoint family, resolution, GPU, and pre-culling design.

Checkpoint	Mean Gaussians	Base (ms)	Policy	Pre-cull	Both	Gain from stacking
Legacy dense A	2,933,368	3.63	$1.080\times$	$1.056\times$	$\mathbf{1.142\times}$	$1.058\times$
Legacy dense B	2,519,315	2.52	$1.138\times$	$1.080\times$	$\mathbf{1.235\times}$	$1.085\times$
DashGaussian	2,167,600	2.47	$1.144\times$	$1.063\times$	$\mathbf{1.220\times}$	$1.067\times$
Speedy-Splat	407,593	0.78	$1.132\times$	$0.988\times$	$\mathbf{1.115\times}$	$0.984\times$

Appendix

Table 13: Policy truncation primarily reduces pair-correlated work; image-bit-exact region pre-culling skips conservatively rejected Gaussian blocks before projection. Standard dataset resolution, planning cost included. This five-arm stress microbenchmark uses at most twelve evenly spaced test views per scene, so its policy column is not a replacement for the all-view twelve-cycle main-table timing. The pre-culling arm’s maximum absolute pixel difference against the unmodified render is

0

in all

52

runs, so the quality cost of the combination is that of the policy alone. Legacy dense A/B are historical packaged checkpoints used only as density stress tests, not as evidence for a named model family. Below roughly

500

K Gaussians the frustum test costs more than it saves and the dispatch rule of Section B.3 disables it.

Backbone	Exact- $1/255$	Ours ( $K_{\mathrm{reg}}=1$ )	Ours ( $K_{\mathrm{reg}}=64$ )	$K_{\mathrm{reg}}{=}64/K_{\mathrm{reg}}{=}1$
Speedy-Splat	$1.000\times$	$1.048\times$	$1.095\times$	$1.047\times$
2DGS	$1.313\times$	$1.422\times$	$1.474\times$	$1.037\times$
Scaffold-GS	$1.252\times$	$1.301\times$	$1.320\times$	$1.014\times$

Appendix

Table 14: Attribution behind the simplified backbone comparison in Table 1 . Exact-

1/255

replaces a released

3\sigma

tile box by the opacity-aware bound already established by prior work [ Radl et al., 2024 , Hanson et al., 2025a ] ; we do not count that transition as a compiler contribution.

K_{\mathrm{reg}}=1

and

K_{\mathrm{reg}}=64

are our compiler with one and 64 decisions. Values are speedups over the official released path except for the final incremental column.

Checkpoint	Resolution	Dataset	Base $\rightarrow$ +Ours [95% CI] (wins)	AccuTile $\rightarrow$ +Ours	$3\sigma\rightarrow$ AccuTile	$\Delta$ PSNR
3DGS	standard	All 13	$\mathbf{1.493\times}$ $[1.350,1.656]$ (13/13)	$\mathbf{1.027\times}$ $[1.014,1.042]$ (13/13)	$1.454\times$	$-0.003$
	1080p	Mip-NeRF 360	$1.729\times$ $[1.532,1.959]$ (9/9)	$1.042\times$ $[1.023,1.064]$ (9/9)	$1.659\times$	$-0.014$
		Tanks&Temples	$2.246\times$ $[2.013,2.506]$ (2/2)	$1.027\times$ $[1.020,1.035]$ (2/2)	$2.186\times$	$-0.009$
		Deep Blending	$2.003\times$ $[1.947,2.060]$ (2/2)	$1.028\times$ $[1.021,1.036]$ (2/2)	$1.948\times$	$+0.003$
		All 13	$\mathbf{1.841\times}$ $[1.657,2.035]$ (13/13)	$\mathbf{1.038\times}$ $[1.024,1.054]$ (13/13)	$1.774\times$	$-0.010$
Mini-Splatting	standard	All 13	$\mathbf{1.346\times}$ $[1.260,1.451]$ (13/13)	$\mathbf{1.063\times}$ $[1.046,1.081]$ (13/13)	$1.266\times$	$-0.008$

Appendix

Table 15: TileSkipper on the two checkpoint families used by Fragment Pruning. Standard-resolution and fixed

1920{\times}1080

rendering on RTX 5000 Ada; each row uses the same patched vanilla-3DGS CUDA rasterizer and twelve balanced timing cycles. “Base

\rightarrow

+Ours” is the deployment-facing total; “AccuTile

\rightarrow

+Ours” isolates our compiled policy after the prior

3\sigma\rightarrow 1/255

exact-bound transition. Intervals bootstrap scenes and parentheses count wins. These are not head-to-head Jetson/OpenGL measurements against Fragment Pruning.

Split seed	$\Delta$ PSNR	$R$	vs. AccuTile	Wins	Compile (s)
17	$-0.0097$	$0.8325$	$1.0934\times$ $[1.0709,1.1176]$	$13/13$	$9.1$
43	$-0.0073$	$0.8283$	$1.0947\times$ $[1.0752,1.1160]$	$13/13$	$9.1$
89	$-0.0084$	$0.8288$	$1.0953\times$ $[1.0762,1.1178]$	$13/13$	$9.1$

Appendix

Table 16: Calibration/selection split stability on all thirteen Speedy-Splat scenes. Each row independently samples 16 calibration and 32 disjoint selection views;

K_{\mathrm{reg}}=64

and all compiler settings are fixed. Speedup intervals bootstrap scenes.

Validation p99 cap	Test p99 $>0.012$	Test p99 $\geq 0.015$	Worst test p99	Mean $R$	Mean worst-view teacher PSNR
$0.012$	$18/39$	$8/39$	$0.0320$	$0.830$	$52.06$ dB
$0.010$	$5/39$	$5/39$	$0.0305$	$0.851$	$54.08$ dB
$0.008$	$3/39$	$3/39$	$0.0301$	$0.877$	$56.43$ dB

Appendix

Table 17: Validation-cap guard-band diagnostic over the same 39 randomized split–scene candidate tables. We reselect policies using held-out training-view measurements alone, then evaluate them on official test views. The original

0.012

row reproduces the reported policies. Tightening the validation cap reduces unseen-view violations at the cost of retaining more pairs, while kitchen failures persist. This post-hoc diagnostic is not used to select the paper’s primary operating point.

Figure 7: Largest safe global tile threshold varies by $16\times$ across regions of a single view. Each grid cell of a representative playroom render is colored by the largest $\tau^{\star}\in\{1,2,4,8,16\}/255$ whose local PSNR stays within $0.03$ dB of the $1/255$ baseline. The four highlighted examples are at (1) $16/255$ (low detail), (2) $8/255$ (material patch), (3) $4/255$ (fine detail), and (4) $2/255$ (edge or boundary). Other grid cells include the conservative $1/255$ value, giving the reported full spread. This is a local tolerance diagnostic rather than an evaluation of the compiled partition.

Figure 8: Grid-aligned residual diagnostic. The crop is selected by maximum boundary-weighted residual energy and the signed residual is amplified $20\times$ ; white lines show the true tile grid. The phase plot pools the Regional policy over three scenes and 92 test views. The highlighted phase zero is the actual tile border. The plot diagnoses residual structure and is not a perceptual visibility threshold.

Figure 9: Backbone versus the compiled policy on held-out kitchen views. Crops are placed automatically at the window of largest disagreement, not chosen by hand; the difference column is amplified $20\times$ . PSNR is against ground truth, so it is comparable to Table 1 .

Explore similar work

Apr 21, 2026cs.CV

AdaGScale: Viewpoint-Adaptive Gaussian Scaling in 3D Gaussian Splatting to Reduce Gaussian-Tile Pairs

Reducing the number of Gaussian-tile pairs is one of the most promising approaches to improve 3D Gaussian Splatting (3D-GS) rendering speed on GPUs. However, the importance difference existing among Gaussian-tile pairs has never been considered in the previous works. In this paper, we propose AdaGScale, a novel viewpoint-adaptive Gaussian scaling technique for reducing the number of Gaussian-tile pairs. AdaGScale is based on the observation that the peripheral tiles located far from Gaussian center contribute negligibly to pixel color accumulation. This suggests an opportunity for reducing the number of Gaussian-tile pairs based on color contribution. AdaGScale efficiently estimates the color contribution in the peripheral region of each Gaussian during a preprocessing stage and adaptively scales its size based on the peripheral score. As a result, Gaussians with lower importance intersect with fewer tiles during the intersection test, which improves rendering speed while maintaining image quality. The adjusted size is used only for tile intersection test, and the original size is retained during color accumulation to preserve visual fidelity. Experimental results show that AdaGScale achieves a geometric mean speedup of 13.8x over original 3D-GS on a GPU, with only about 0.5 dB degradation in PSNR on city-scale scenes.

Joongho Jo, Hyerin Lim, Hanjun Choi +1

Korea University Seoul, Republic of Korea

Cpu-Gpu Hybrid Designs Distortion

Ask PDF

May 29, 2026cs.CV

HiGS: A Hierarchical Rendering Architecture for Real-Time 3D Gaussian Splatting

3D Gaussian Splatting (3DGS) has become the standard for real-time novel view synthesis on commodity GPUs. Its pipeline ties spatial partitioning and rasterization to one tile size, yet the two pull in opposite directions: partitioning, which bins and depth-sorts gaussians, grows cheaper with larger tiles, while rasterization gets cheaper with smaller ones. Prior acceleration work reduces the cost of individual stages but keeps both locked to that single scale, where a few dense tiles dominate frame time. We present Hierarchically Tiled Gaussian Splatting (HiGS), which gives each its own scale: partitioning runs over coarse macro-tiles, while rasterization runs over the fine render tiles within them. Rasterization work is then issued in proportion to the gaussians in each macro-tile rather than per tile, so dense regions spread across many parallel units instead of serializing through one. Across tested scenes, HiGS renders up to 15.8x faster than the original 3DGS and outperforms every other rasterizer we evaluate, while preserving exact front-to-back alpha compositing.

Dawid Pająk, Martin Bisson, Rodolfo Lima

NVIDIA

Rasterization Gaussian Splatting Partitioning Hierarchical

Ask PDF

May 6, 2026cs.CV

QuadBox: Accelerating 3D Gaussian Splatting with Geometry-Aware Boxes

3D Gaussian Splatting (3DGS) has emerged as an advanced technique for real-time novel view synthesis by representing scene geometry and appearance using differentiable Gaussian primitives. However, efficiently computing precise Gaussian-tile intersections remains a critical task in the rasterization pipeline. To this end, we propose QuadBox, a method that leverages four axis-aligned bounding boxes to tightly encapsulate projected Gaussians in a discrete manner. First, we derive a geometry-aware stretching factor that enables the construction of a tile-aligned QuadBox, which covers the elliptical projection and largely excludes irrelevant tiles. Second, we introduce QPass, a single-pass tile traversal algorithm that exhaustively exploits the discrete nature of QuadBox, ensuring that the tile intersection check is performed with simple interval tests. Experiments on public datasets show that our method accelerates the rendering speed of 3DGS by 1.85

\times

. Code is available at https://github.com/Powertony102/QuadBox.

Xinze Li, Bohan Yang, Pengxu Chen +4

Beijing Normal-Hong Kong Baptist University · Jilin University · Hong Kong Baptist University +2

3D Gaussian Rasterization Bounding Box

Ask PDF