Tier A complete: 25/25 runs HiGS accel gate PASS · 1.164x training speedup at matched quality

3DGS Renderer Benchmark and Research Suite

Three evidence-separated tracks: reproducible method survey, differentiable HiGS training, and storage compression. Published rankings remain limited to their declared A100 cohort.

3DGS Renderer Benchmark preview Star on GitHub
1.164xHiGS training speedup · formal gate PASS
132Confirmatory jobs (11 scenes x 3 seeds)
696Fastest FPS (HiGS)
5.73xSPZ near-lossless cohort
25/25Accepted runs
5Scenes measured

New here? Start with the quickstart

1 clone & install · 2 prepare the pinned Garden case · 3 run an installed backend and inspect results/. 
git clone https://github.com/caizefan34/3dgs-renderer-benchmark.git
cd 3dgs-renderer-benchmark
pip install -r requirements-benchmark.txt
python benchmark.py prepare mipnerf360 --scene garden
python benchmark.py prepare-case small-garden-1080p
python benchmark.py run gsplat_higs --dataset garden
READMEFull project guide: quickstart, FAQ, repository layout Documentation guideWhere to find everything in the repo Research programThree tracks, evidence contract, and publication boundaries Differentiable HiGSCurrent evidence and confirmatory speedup gates MethodologyHow the numbers were measured (evidence tiers)

Measured renderer cohort

Choose gsplat HiGS for maximum throughput (696 FPS aggregate), Speedy-Splat for reference-level quality and balanced deployment, gsplat for the lowest peak memory (4,206 MiB), and Original 3DGS when an official reference path matters most.

Three research tracks

Track 1 · Quality-preserving speedup PASS

Differentiable HiGS

Native CUDA backward, frozen and dynamic topology paths, gradient checks, and a frozen 132-job 3-seed confirmatory matrix. Candidate gsplat_30k_fused_prune10_rclip05 passes all pre-registered gates: PSNR CI lo -0.022, SSIM CI lo -0.0011, LPIPS CI hi +0.0025, 1.164x wall speedup (CI lo 1.034), TTQ faster. Quality preserved with >=10% faster training, not ordinary early-stop (gsplat_25k fails the quality gates). Open paper plan →

Track 2 · Search audit blocked

Reproducible 3DGS survey

The registry pins sources and separates reported, reproduced, comparable, and blocked evidence. Exhaustive or latest-survey language requires a frozen database search and second-reviewer audit. Open survey protocol →

Track 3 · Five-scene qualification

Lossless and near-lossless storage

SPZ 8/8 passes the declared same-checkpoint near-lossless gate on five scenes at 5.572x to 6.072x compression. Bit-exact XZ and retraining-required codecs remain separate categories. Open compression protocol →

Renderer rankings

Below: 25 accepted runs (5 renderers × 5 scenes). Colors stay consistent. Bars show within-cohort rank with exact measured values.

FPS

Tier A aggregate FPS ranking

PSNR

Tier A aggregate PSNR ranking

SSIM

Tier A aggregate SSIM ranking

LPIPS

Tier A aggregate LPIPS ranking

VRAM

Tier A peak process VRAM ranking

Speed vs Quality

FPS versus LPIPS Pareto chart

Overall comparison table

Chart data: 25 accepted runs (5 renderers × 5 cases), within one EPIC-05 A100 Tier A cohort.

RendererSpeed indexFPSPSNRSSIMLPIPSPeak VRAM
gsplat HiGS5.671×696.9125.8340.83720.26496,616 MiB
Speedy-Splat2.385×293.0326.1210.84040.26374,276 MiB
TC-GS2.048×251.6226.1300.84010.26334,322 MiB
gsplat1.966×241.6025.8340.83720.26534,206 MiB
Original 3DGS1.000×122.8826.1200.84030.26378,234 MiB

Explore the evidence

Comparison analysisTrade-offs, case winners, limitations, decision guide Ranking tablesOverall, speed, quality, memory, efficiency, Pareto Machine-readable dataMetrics, confidence intervals, cohorts, recommendations MethodologyEvidence tiers, canonical workload, measurement rules Documentation guideIntegration references, historical context HiGS paper planMethod scope, current support, and confirmatory gate results Final conclusionsComprehensive data analysis, winners, future directions Run reportsMachine cohort, dataset hashes, archived Windows audit

Support this project

If this benchmark helps your research or engineering decisions,
star the repo on GitHub →