PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.09343v1cs.CV

TL;DR

Tiled 3D Gaussian Splatting uses scene-wide contribution cutoffs even though content varies in its sensitivity to truncation. TileSkipper compiles static cutoffs for frozen checkpoints using measured pair savings and compositing risk, then validates complete renders. Across six integrations with opacity-aware bounds, compiler-only speedups range from 1.009× to 1.121×.

  • Problem

    A scene-wide tile cutoff cannot allocate error differently across content with differing sensitivity to support truncation.

    Source: §1 · p. 2

  • Method

    TileSkipper measures pair savings and compositing-aware isolated-removal risk, allocates static cutoff levels across regions, and validates proposals with complete renders.

    Source: §1 · p. 2, §3 · p. 5

  • Results

    Compiler-only speedups are 1.096× on Speedy-Splat, 1.009× on FastGS, 1.091× on SeeLe, 1.121× on FlashGS, 1.095× on LightGaussian, and 1.051× on Taming-3DGS.

    Source: §4.2 · p. 8

  • Takeaways & Limitations

    Across ten model/backend integrations, a one-byte-per-Gaussian static policy requires no training or per-frame policy inference and composes with compact checkpoints and independently optimized rasterizers.

    Source: §5 · p. 10

    The isolated-deletion risk proxy omits cross terms when many contributions are removed jointly, and held-out selection does not guarantee test-view residuals stay within the cap.

    Source: §3.2 · p. 6, §D.10 · p. 22

Abstract

from arXiv · show

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 $3\sigma$ 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.

1 INTRODUCTION

TileSkipper targets excess Gaussian–tile work from a scene-wide cutoff that cannot adapt to differing content sensitivity. It compiles a static, heterogeneous cutoff policy for frozen checkpoints and validates proposals with complete renders.

1 INTRODUCTION

  • Because per-frame cost scales with Gaussian–tile pairs, large projected footprints increase binning, sorting, and rasterization work.Reducing this work matters particularly for stereo rendering and mobile or XR hardware with limited throughput.Source: §1 · p. 1
  • A scene-global threshold must compromise between content that tolerates aggressive skipping and edges, fine textures, or highlights that require conservative cutoffs.One playroom view has locally safe thresholds spanning the full tested range from τ⋆=1/255 to 16/255 across regions.Source: §1 · pp. 1–2
  • At matched teacher PSNR, Regional is 1.040–1.046× faster than OURS-GLOBAL at standard resolution and 1.041–1.083× faster at 4K, tying the per-Gaussian variant.On the diagnostic work–fidelity frontier, Regional reaches 16.4 dB higher teacher PSNR than the swept scene-global cutoff at the same pair count; this is not a ground-truth PSNR gain.Source: Figure 1 · p. 2
  • Unlike AdaGScale’s runtime view adaptation, TileSkipper measures compositing residuals over calibration views and uses a static scene-level policy without per-frame inference.The checkpoint-matched standalone comparison is dataset-dependent and does not establish broad runtime dominance.Source: §1 · p. 2
  • TileSkipper measures candidate pair savings and compositing-aware isolated-removal residuals, then allocates cutoff levels across risk-stratified groups and validates proposals with complete renders.The exported policy is one uint8 level per Gaussian, with no checkpoint update or per-frame policy inference.Source: §1 · p. 2
  • The post-training policy is static and portable, and policies compiled on AccuTile transfer unchanged into SeeLe and FlashGS.A measured cost model separates pair-dependent work the policy can reduce from fixed work it cannot, and predicts when complementary pre-culling pays.Source: §1 · p. 3

2 RELATED WORK

TileSkipper operates on tile-pair enumeration, building on tighter opacity-aware support bounds while selecting heterogeneous lossy cutoffs through measured compositing risk. Its static policy differs from runtime-adaptive support and from methods that prune at other pipeline stages.

2 RELATED WORK

  • Unlike Fragment Pruning’s learned per-Gaussian fragment-level thresholds after sorting, TileSkipper selects discrete tile-enumeration cutoffs before sorting without updating checkpoint parameters.The checkpoint-family evaluation tests compatibility, not composition with learned fragment pruning.Source: §2 · p. 3
  • The SeeLe transfer changes its fixed tile cutoff while retaining the system’s other components, rather than reproducing its full mobile deployment pipeline.SeeLe combines camera-cluster preprocessing with contribution-aware rasterization.Source: §2 · p. 3
  • TileSkipper borrows prior opacity-level-set bounds rather than claiming the support formula as new; heterogeneous lossy cutoff selection is complementary to exact or fixed-threshold bounds.AdR-Gaussian, Speedy-Splat, FastGS, FlashGS, and QuadBox provide related tile-support constructions.Source: §2 · p. 3
  • TileSkipper changes stage 2, tile enumeration: excluded tiles create no Gaussian–tile pair and therefore enter neither sorting nor blending.The figure contrasts baseline and compiled support and shows the resulting depth order and compositing expression.Source: Figure 2 · p. 4
  • AdaGScale adapts tile support to the current view, whereas TileSkipper measures compositing residuals over calibration views and validates a static policy with complete renders.TileSkipper trades runtime view adaptation for no per-frame policy inference; its standalone comparison is inside AdaGScale’s released renderer.Source: §2 · p. 4

3 METHOD

TileSkipper replaces a scene-wide tile cutoff with a static per-Gaussian policy for a frozen checkpoint. It estimates pair savings and compositing risk, selects and validates regional cutoff levels, then applies the exported policy during tile enumeration.

3 METHOD

  • The policy replaces one scene-wide cutoff with static per-Gaussian multipliers; Kreg is the number of independently selected regions, and mg = 1 recovers Official-Fixed.OURS-GLOBAL uses one scene-specific multiplier, while OURS-REGIONAL selects levels for Kreg = 64 regions.Source: §3 · p. 4
  • TileSkipper changes only tile enumeration, leaving Gaussian parameters, depth sorting, and blending arithmetic unchanged.At a fixed checkpoint and resolution, latency is approximated as T ≈ F + βP, where F is pair-independent cost and P is Gaussian–tile pair count.Source: §3 · p. 4

3.1 TILE SUPPORT FROM A SCENE-WIDE CUTOFF

  • The τ-level opacity support is prior work; raising τ shrinks the footprint and removes pairs, but individually small discarded contributions can accumulate into image error.This motivates selecting heterogeneous τg values from measured rendering risk and validating complete candidate renders.Source: §3.1 · p. 5

3.2 COMPILING A DISCRETE TAIL POLICY

  • For each Gaussian and cutoff rung, calibration measures exact pair savings and an isolated-removal RGB-energy proxy that accounts for compositing sensitivity.The proxy omits cross terms when many contributions are removed jointly, so it ranks candidates rather than exactly predicting joint image error.Source: §3.2 · p. 6
  • Eight efficiency-quantile strata are partitioned by Morton order into workload-balanced regions; work balancing reduces region-work coefficient of variation from 3.1–4.8 to 0.026–0.060.A constrained search proposes regional levels and includes uniform-level candidates so the regional policy can fall back to a global cutoff.Source: §3.2 · p. 6
  • Disjoint held-out training views screen each candidate by minimum teacher PSNR and worst-view p99 residual, and the passing policy with greatest pair savings is exported.Runtime reads one uint8 level per Gaussian and a lookup table inside tile enumeration, with no extra kernel or per-frame policy inference.Source: §3 · p. 5, §3.2 · p. 7

4 EXPERIMENTS

The experiments evaluate TileSkipper across model/backend integrations and resolutions, then compare regional allocation with global and per-Gaussian control and assess it against AdaGScale. Results distinguish compiler-only gains from prior exact-bound transitions and show that regional allocation improves on global calibration while tying per-Gaussian control.

4 EXPERIMENTS

  • The experiments assess ten model/backend integrations, including unchanged policy transfer to SeeLe and FlashGS and checkpoint variations on a common AccuTile backend.The table’s comparisons are base model/backend versus the same path plus TileSkipper; some rows include a prior exact-bound transition.Source: §4 · p. 7, Table 1 · p. 8

4.1 SETUP

  • Evaluation covers thirteen scenes, paired same-process timing, standard metrics, and a 4K setting; calibration and selection use disjoint training views, not test views.The scenes span Mip-NeRF 360, Tanks & Temples, and Deep Blending; standard-resolution images range from 0.53–1.62 megapixels.Source: §4.1 · pp. 7–8

4.2 MAIN RESULTS

  • 1.009×–1.121× compiler-only speedups appear on six paths with existing opacity-aware bounds; other ports separately include prior exact-bound transitions.For vanilla 3DGS, the total 1.493× factor decomposes into a 1.454× exact-bound transition and a 1.027× compiler gain; 2DGS and Scaffold-GS improve by 1.474× and 1.320×, respectively.Source: §4.2 · p. 8
  • The measured gain grows with resolution, while fixed per-frame MLP work limits the Scaffold-GS end-to-end improvement.Duplicate-plus-sort shares rise from 21.4% to 34.2% on bicycle and from 31.6% to 41.6% on counter.Source: §4.2 · p. 8

4.3 WHAT REGIONAL ALLOCATION ADDS

  • Regional allocation is 4.0–4.6% faster than global at native resolution and up to 8.3% faster at 4K, but ties per-Gaussian control at 0.986–1.007×.Both regional and per-Gaussian policies export one byte per Gaussian, so the comparison shows no demonstrated regional storage advantage.Source: §4.3 · p. 9, Table 2 · p. 9

4.4 CLOSEST-METHOD COMPARISON

  • At native resolution, TileSkipper’s pooled speedup over AdaGScale is 1.010× [0.990, 1.025]; at 4K it is 1.076× [1.047, 1.102], with 12/13 scene wins.These standalone comparisons use common training-view quality constraints; the native interval does not establish broad runtime dominance.Source: §4.4 · p. 9, Table 3 · p. 9

5 CONCLUSION

TileSkipper compiles a static heterogeneous cutoff policy for multiple model/backend paths without training or per-frame policy inference.

5 CONCLUSION

  • Across ten model/backend integrations, TileSkipper compiles a scene-wide tile-footprint threshold into a heterogeneous static policy that composes with compact checkpoints and independently optimized rasterizers.The one-byte-per-Gaussian policy requires no training or per-frame inference.Source: §5 · p. 10

REPRODUCIBILITY STATEMENT

The paper describes its method, experiments, and supporting appendices as resources for reproducing the reported results.

REPRODUCIBILITY STATEMENT

  • The method, evaluation protocols, backend integration details, threshold semantics, and experimental protocols are described to support reproduction of the results.The authors state that source code and evaluation scripts will be publicly released upon acceptance.Source: REPRODUCIBILITY STATEMENT · p. 10

AI USE STATEMENT

The paper reports language-polishing assistance and separately positions TileSkipper by rasterization stage and decision unit, with optional pre-culling distinct from its core threshold policy.

AI USE STATEMENT

  • Generative AI tools assisted with language polishing, and the authors take responsibility for the paper’s content.Source: AI USE STATEMENT · p. 10

A RELATED-WORK COMPARISON

  • TileSkipper uses heterogeneous static per-Gaussian cutoffs without checkpoint updates or a hand-set scene-wide cutoff, validating proposed policies with complete held-out renders.Source: Table 4 · p. 13

B COST MODEL AND OPTIONAL PRE-CULLING

B.1 A THREE-TERM COST MODEL

The cost model separates pair-dependent latency from fixed work and uses measured savings to decide when pre-culling pays. Pre-culling is image-bit-exact in the evaluated runs, while its savings are not assumed to add to thresholding savings.

B.1 A THREE-TERM COST MODEL

  • R2 = 0.9934–0.9988 across nine diagnostic cells, supporting pair count as a latency surrogate within the measured fixed-policy-family cutoff sweeps.The fit does not establish equal per-pair cost or sufficiency across policy families, checkpoint densities, or pre-culling configurations.Source: §B.1 · p. 13
  • The fitted pair-independent preprocessing intercept is 35.1% of baseline frame latency on bicycle, a component threshold policies cannot reduce.The fitted slope was −0.023 ns per pair with R2 = 0.26; the low R2 reflects only 26% of this stage’s small timing variation, not a test that the slope is zero.Source: §B.1 · p. 13
  • Duplicate-plus-sort share rises with resolution from 21.4% to 34.2% for bicycle and 31.6% to 41.6% for counter.The estimated perfect, free visibility-culler ceiling falls from 22.6% to 7.6% for bicycle and from 9.3% to 2.2% for counter.Source: §B.1 · p. 13, Figure 4 · p. 14

B.2 CONSERVATIVE REGION PRE-CULLING

  • Pre-culling skips conservatively rejected Gaussian blocks before projection, and its images were bit-identical to the unmodified backbone in all 52 evaluated runs.Planning time is included in the reported pre-culling timing; the authors limit the claim to image-bit-exact behavior in evaluated runs, not arbitrary cameras or hardware.Source: §B.2 · p. 14

B.3 DISPATCH: WHICH MECHANISM, AND WHEN

  • Across thirteen scenes, the empirical dispatch model estimates the realized pre-culling ceiling with mean absolute error 0.004, without treating the mechanisms’ savings as additive.The estimate is derived from one oracle visible-subset comparison and is scoped to the tested checkpoint, resolution, and hardware.Source: §B.3 · p. 14
  • At standard resolution, aN/T > 20% is a reliable pre-culling “on” regime and values below 10% are reliably “off”; intermediate cases use measured net benefit over planning overhead.Dispatch is not based on Gaussian count alone.Source: §B.3 · p. 14

C ADDITIONAL COMPARISON PROTOCOLS AND RESOLUTION RESULTS

TileSkipper’s regional policies improve on scene-global calibration, while the experiments probe resolution scaling, policy granularity, construction choices, temporal behavior, transfer, and pre-culling dispatch. The reported gains depend on the comparison regime and workload; several ablations identify boundaries where extra complexity does not establish additional benefit.

C.1 COMPILER GRANULARITY PROTOCOL

  • Regional allocation beats scene-global calibration by 4.0–4.6% at standard resolution and 8.3% at 4K, while tying per-Gaussian control at 0.986–1.007×.The comparison interpolates policy curves to common teacher-PSNR targets; both regional and per-Gaussian implementations export one level byte per Gaussian.Source: §C.1 · p. 15

C.2 ADAGSCALE COMPARISON PROTOCOL

  • In Table 3, AdaGScale recalibrates its twenty-bin Tupper table per checkpoint and resolution, while TileSkipper re-evaluates a frozen native-resolution Kreg = 64 family; shared selection constraints do not enforce equal attained quality.Source: §C.2 · p. 15

D.1 FIXED-POLICY RESOLUTION STRESS TEST

  • Dataset-macro speedups are 1.088× at standard resolution and 1.238× at 3840-pixel width for unchanged policies selected at standard resolution.The higher-resolution experiment resamples the same source views and therefore tests workload scaling, not newly captured image detail.Source: §D.1 · p. 15, Table 5 · p. 16

D.2 REGION-COUNT ABLATION ON SPEEDY-SPLAT; D.3 REGION-COUNT ABLATION ON 2DGS

  • On Speedy-Splat, Kreg = 64 gives 1.0472× over its same-process global control; on 2DGS, Kreg = 64 and 256 reach 1.1216× and 1.1228× over the exact-bound baseline.The 2DGS paired 256-versus-64 ratio is 1.0011× [0.9988, 1.0037], so the sweep establishes no additional runtime benefit beyond 64 groups.Source: §D.2 · p. 16, Table 6 · p. 17, §D.3 · p. 17

D.4 REGION-CONSTRUCTION ABLATION

  • At 55 dB teacher PSNR, the full risk-aware construction reaches 1.103×, versus 1.065× for random groups and 1.066× for a uniform grid.Workload balancing cuts region-work CV from 0.684 to 0.047, but count-balanced risk strata and the full method have statistically indistinguishable latency.Source: §D.4 · p. 17, Table 8 · p. 18

D.5 WHY NOT SIMPLY RAISE THE CONSTANT?

  • A fixed cutoff matched to regional worst-scene quality reaches only 1.036×, while matching the regional 1.1009× speed requires a fixed cutoff whose worst scene loses 0.334 dB.The selector’s per-scene optimum ranges from 1.5–3.0τ0 and requires scene measurements and rendered acceptance tests.Source: §D.5 · pp. 18–19, Table 9 · p. 19

D.6 DENSE, FLOW-ALIGNED TEMPORAL EVALUATION

  • On three scenes and 359 interpolated frames, regional adaptation reduces aligned residual change to 0.789× the global value, but normalized aggregate change is 0.990×, supporting neither regression nor improvement.Counter has a 1.050× normalized result, so these paths do not establish a universal no-popping guarantee.Source: §D.6 · pp. 19–20, Table 11 · p. 20

D.7 TRANSFER TO INDEPENDENTLY OPTIMIZED RASTERIZERS; D.8 THE COST MODEL PREDICTS, AND ONE OF ITS PREDICTIONS IS NEGATIVE; D.9 COST

  • Unchanged policies transfer to SeeLe at 1.091×/1.260× and FlashGS at 1.121×/1.317× at standard/4K resolution; pre-culling instead yields 1.010× overall and about 1.235× combined on tested checkpoints above ∼2M Gaussians.Below ∼500K Gaussians, pre-culling reaches 0.988× on Speedy-Splat and dispatch switches it off; default compilation takes 9.0 seconds per scene and exports N bytes of indices.Source: §D.7 · p. 20, §D.8 · pp. 20–21, Table 12 · p. 20

D.10 LIMITATIONS

The limitations cover calibration requirements, supported rasterizer scope, deployment behavior, and evaluation boundaries. The backbone-transition analysis also separates prior exact-bound gains from TileSkipper’s compiler contribution.

D.10 LIMITATIONS

  • Compilation requires each scene’s training poses and images, but not optimizer state or parameter updates.It is a per-scene calibration step, not a data-free transform of an arbitrary released .ply.Source: §D.10 · p. 22
  • The method applies to threshold-based tile enumeration; oriented-box, convex-hull, and non-tiled rasterizers expose different controls and are out of scope.The stated direct scope includes AccuTile-style rasterizers, SeeLe, and other listed rasterizers once their 3σ box is replaced by a contribution-threshold bound.Source: §D.10 · p. 22
  • Regional thresholds tie per-Gaussian control on speed, and both policies use one byte per Gaussian, so regional control is not a compression claim.Regions reduce independent search decisions and may regularize selection, but the current packed format stores the same number of bytes per Gaussian.Source: §D.10 · p. 22
  • The exported policy is static; the 359-frame, three-scene flow-aligned test had no aggregate regression, but counter regressed 5.0% after normalization.Broader trajectories and a perceptual user study are needed for a universal no-popping claim.Source: §D.10 · p. 22
  • The 0.012 p99 cap applies to held-out training views, not every unseen pose: 18 of 39 split–scene runs exceeded it, with a maximum of 0.032.The diagnostics use floating-point residuals and the same 4× spatial subsample, so the excesses are not a dtype mismatch.Source: §D.10 · p. 22
  • Tile-support truncation creates tile-aligned residual structure, but the measured boundary signal does not establish perceptual visibility.Across three scenes and 92 test views, boundary/non-boundary ratios were 5.86–6.18×; a controlled user study remains absent.Source: §D.10 · p. 22
  • All timings use RTX 5000 Ada GPUs; a second desktop or mobile hardware class remains unmeasured.The cost model suggests the balance between the two mechanisms may shift on hardware with a different memory-to-compute ratio.Source: §D.10 · p. 22

E BACKBONE-TRANSITION ATTRIBUTION

  • Speedy-Splat’s full gain is compiler-driven, while 2DGS and Scaffold-GS first gain speed from the prior 3σ-to-1/255 exact-bound transition.The deployment-facing comparison combines these stages for the latter ports, so their total speedups should not be attributed entirely to the compiler.Source: E BACKBONE-TRANSITION ATTRIBUTION · p. 22

F THE CHECKPOINT FAMILIES USED BY FRAGMENT PRUNING

TileSkipper was tested independently on vanilla 3DGS and Mini-Splatting checkpoints, not as Fragment Pruning or as a cross-hardware speed contest. The reported totals include prior exact-bound gains, and unseen-view quality can exceed selection constraints.

F THE CHECKPOINT FAMILIES USED BY FRAGMENT PRUNING

  • The experiment tests compatibility with the checkpoint families, not superiority to Fragment Pruning; different hardware, renderer, and FPS cap make direct ranking invalid.TileSkipper uses an RTX 5000 Ada CUDA rasterizer, whereas Fragment Pruning’s published FPS uses a Jetson Orin NX and OpenGL.Source: F THE CHECKPOINT FAMILIES USED BY… · p. 23
  • 1.493× and 1.346× are the standard-resolution deployment totals for 3DGS and Mini-Splatting; compiler-only regional gains are 1.027× and 1.063×.At fixed 1080p, totals are 1.841× and 1.599×, with compiler-only gains of 1.038× and 1.089×; all thirteen scenes were faster in both settings.Source: F THE CHECKPOINT FAMILIES USED BY… · p. 23
  • On unseen test views, mean ground-truth PSNR changes were −0.010/−0.014 dB for 3DGS/Mini-Splatting, while Mini-Splatting’s worst fixed-1080p view reached 0.0259 p99.The selected policies passed 48-dB and 0.012 p99 constraints on held-out training views; those constraints do not certify unseen poses.Source: F THE CHECKPOINT FAMILIES USED BY… · p. 23

G CALIBRATION-VIEW SPLIT STABILITY

Random calibration/selection splits produce stable aggregate speedups despite finding different near-equivalent policies. Tighter validation caps reduce unseen-view violations at a work cost, while local safe thresholds vary across a view.

G CALIBRATION-VIEW SPLIT STABILITY

  • 0.0010× was the standard deviation in thirteen-scene speedup across three random splits, and all three improved all thirteen scenes.Per-Gaussian multiplier vectors agreed exactly only 49.5% of the time; mean compilation time was 9.1 seconds per scene.Source: G CALIBRATION-VIEW SPLIT STABILITY · p. 23, Table 16 · p. 24
  • Tightening the validation p99 cap from 0.012 to 0.010 and 0.008 reduced unseen-view violations from 18 to 5 and 3 of 39 runs, while retained-pair ratio rose to 0.877.All three failures at 0.008 were kitchen, where worst p99 remained 0.021–0.030; a global margin did not address view coverage.Source: G CALIBRATION-VIEW SPLIT STABILITY · p. 24

H SPATIAL HETEROGENEITY OF SAFE GLOBAL THRESHOLDS

  • A single playroom view contains a 16× spread in the largest locally safe global threshold, from 1/255 to 16/255.This pixel-local diagnostic establishes heterogeneity, not deployable regional gain; playroom is one of two scenes whose selected Regional policy collapses to one level.Source: H SPATIAL HETEROGENEITY OF SAFE GLOBAL… · p. 24

I TILE-BOUNDARY RESIDUAL STRUCTURE

Tile-level support truncation produces residual structure aligned with tile boundaries, although this diagnostic does not establish perceptual visibility. The qualitative comparison focuses on the largest-disagreement crops in held-out kitchen views.

I TILE-BOUNDARY RESIDUAL STRUCTURE

  • The local safe-threshold diagnostic spans 1/255 to 16/255 across one view, but it is not an evaluation of the compiled partition.The four highlighted regions correspond to low detail at 16/255, a material patch at 8/255, fine detail at 4/255, and an edge or boundary at 2/255.Source: Figure 7 · p. 25
  • 5.86×, 6.18×, and 5.93× are the boundary/non-boundary residual ratios for Global, Regional, and per-Gaussian control, respectively.The true tile boundary was the largest of 16 phases in all 92 views for every policy; mean boundary jumps were 0.192, 0.287, and 0.312 of one 8-bit code value.Source: I TILE-BOUNDARY RESIDUAL STRUCTURE · p. 25, Figure 8 · p. 25

J QUALITATIVE COMPARISON

  • Held-out kitchen crops show +0.00 to −0.02 dB per-view PSNR changes; the largest-disagreement windows are amplified 20× and are not representative crops.The largest single-pixel difference was 0.074 on a [0, 1] scale, concentrated on the woven placemat and specular window-frame edge.Source: J QUALITATIVE COMPARISON · p. 26, Figure 9 · p. 26
Loading 2610.09343v1…