GoDe: Gaussians on Demand

for Progressive Level-of-Detail and Scalable Compression

Francesco Di Sario1,2*
Riccardo Renzulli1
Marco Grangetto1
Akihiro Sugimoto2
Enzo Tartaglione3
1University of Turin
2National Institute of Informatics
3Télécom Paris, Institut Polytechnique de Paris
Università di Torino NII Télécom Paris

*Corresponding author: francesco.disario@unito.it

Paper Code ar Xiv arXiv

Fig. 1: Our method builds a progressive Gaussian hierarchy enabling adaptive levels of detail (LoD). Combined with compression, it supports multiple rates from a single trained model — no retraining required.

Abstract

Recent progress in compressing explicit radiance field representations, and in particular 3D Gaussian Splatting, has substantially reduced memory consumption while further improving real-time rendering performance. Despite these advances, existing approaches are inherently single-rate: each target compression level is obtained through a separately optimized model, resulting in a collection of fixed operating points rather than a truly scalable representation.

We argue that scalable compression should be formulated as an intrinsic property of the representation itself. Building on this observation, we introduce GoDe (Gaussians on Demand), a general framework for scalable compression and progressive level-of-detail control. Starting from a single trained model, GoDe reorganizes Gaussian primitives into a fixed progressive hierarchy that supports multiple discrete rate–distortion operating points without retraining or per-level fine-tuning.

Contributions

🎯
Scalable compression from a single pipeline

All L operating points are derived from one pretrained model after a single joint fine-tuning stage — no separate training per budget required.

📐
Gradient-induced hierarchical organization

An iterative gradient-informed masking strategy aggregates gradient norms across all Gaussian parameters to construct a stable, well-structured LoD hierarchy.

🔧
Model-agnostic design

Works on vanilla 3DGS, Scaffold-GS, and Octree-GS without architectural modifications.

🚀
Practical deployment

Decoding speedups of up to 1762× compared to prior methods via ZSTD progressive coding.

Method

GoDe operates entirely post-training in three stages: (1) gradient-informed iterative masking organizes Gaussians into a progressive hierarchy; (2) a single quantization-aware fine-tuning stage jointly optimizes all levels via random level sampling; (3) each level is independently compressed for scalable progressive decoding.

Fig. 2: Overview of GoDe. (1) Reorganize Gaussians into a progressive hierarchy via iterative gradient-informed assignment, (2) single quantization-aware fine-tuning jointly optimizing all levels, (3) compress each level independently for scalable decoding.

Results

GoDe covers compression rates from a few MB up to ~120 MB while preserving competitive reconstruction quality across all three standard benchmarks and all three 3DGS backbone families. The full rate–distortion curve is produced by a single progressive representation.

Rate-distortion and FPS curves

Fig. 3: (Top) Rate–distortion curves on Mip-NeRF360, Tanks&Temples, and DeepBlending. (Bottom) PSNR / FPS trade-off. Lower LoDs offer 500+ FPS for real-time adaptive rendering.

Progressive Level-of-Detail Visualization

Use the slider to explore all 8 levels of detail for each scene.

Mip-NeRF360

Bicycle

Level of Detail LoD 0
01234567
bicycle LoD 0
bicycle LoD 1
bicycle LoD 2
bicycle LoD 3
bicycle LoD 4
bicycle LoD 5
bicycle LoD 6
bicycle LoD 7
—FPS
—PSNR
—MB
—Gauss.

Bonsai

Level of Detail LoD 0
01234567
bonsai LoD 0
bonsai LoD 1
bonsai LoD 2
bonsai LoD 3
bonsai LoD 4
bonsai LoD 5
bonsai LoD 6
bonsai LoD 7
—FPS
—PSNR
—MB
—Gauss.

Counter

Level of Detail LoD 0
01234567
counter LoD 0
counter LoD 1
counter LoD 2
counter LoD 3
counter LoD 4
counter LoD 5
counter LoD 6
counter LoD 7
—FPS
—PSNR
—MB
—Gauss.

Flowers

Level of Detail LoD 0
01234567
flowers LoD 0
flowers LoD 1
flowers LoD 2
flowers LoD 3
flowers LoD 4
flowers LoD 5
flowers LoD 6
flowers LoD 7
—FPS
—PSNR
—MB
—Gauss.

Garden

Level of Detail LoD 0
01234567
garden LoD 0
garden LoD 1
garden LoD 2
garden LoD 3
garden LoD 4
garden LoD 5
garden LoD 6
garden LoD 7
—FPS
—PSNR
—MB
—Gauss.

Kitchen

Level of Detail LoD 0
01234567
kitchen LoD 0
kitchen LoD 1
kitchen LoD 2
kitchen LoD 3
kitchen LoD 4
kitchen LoD 5
kitchen LoD 6
kitchen LoD 7
—FPS
—PSNR
—MB
—Gauss.

Room

Level of Detail LoD 0
01234567
room LoD 0
room LoD 1
room LoD 2
room LoD 3
room LoD 4
room LoD 5
room LoD 6
room LoD 7
—FPS
—PSNR
—MB
—Gauss.

Stump

Level of Detail LoD 0
01234567
stump LoD 0
stump LoD 1
stump LoD 2
stump LoD 3
stump LoD 4
stump LoD 5
stump LoD 6
stump LoD 7
—FPS
—PSNR
—MB
—Gauss.

Treehill

Level of Detail LoD 0
01234567
treehill LoD 0
treehill LoD 1
treehill LoD 2
treehill LoD 3
treehill LoD 4
treehill LoD 5
treehill LoD 6
treehill LoD 7
—FPS
—PSNR
—MB
—Gauss.
Tanks & Temples

Truck

Level of Detail LoD 0
01234567
truck LoD 0
truck LoD 1
truck LoD 2
truck LoD 3
truck LoD 4
truck LoD 5
truck LoD 6
truck LoD 7
—FPS
—PSNR
—MB
—Gauss.

Train

Level of Detail LoD 0
01234567
train LoD 0
train LoD 1
train LoD 2
train LoD 3
train LoD 4
train LoD 5
train LoD 6
train LoD 7
—FPS
—PSNR
—MB
—Gauss.
Deep Blending

Dr. Johnson

Level of Detail LoD 0
01234567
drjohnson LoD 0
drjohnson LoD 1
drjohnson LoD 2
drjohnson LoD 3
drjohnson LoD 4
drjohnson LoD 5
drjohnson LoD 6
drjohnson LoD 7
—FPS
—PSNR
—MB
—Gauss.

Playroom

Level of Detail LoD 0
01234567
playroom LoD 0
playroom LoD 1
playroom LoD 2
playroom LoD 3
playroom LoD 4
playroom LoD 5
playroom LoD 6
playroom LoD 7
—FPS
—PSNR
—MB
—Gauss.

Smooth Level Transitions

Switching between adjacent levels can cause abrupt pop-in artifacts. GoDe mitigates this via runtime opacity interpolation: enhancement Gaussians in El fade in linearly from opacity 0 to their target value. Applied at render time only — no training overhead.

Garden

Kitchen

Truck

Counter

Bicycle

Bonsai

Train

Playroom

Quantitative Comparison

Comparison against recent compression methods on three standard benchmarks, all retrained and evaluated on a single NVIDIA A40. GoDe presents all 8 levels from a single trained model. Red = best, yellow = second best. ✓ = scalable (single model, multiple rates).

Method Level Scalable Mip-NeRF360
PSNR ↑ SSIM ↑ LPIPS ↓ Size ↓ W ↓ FPS ↑
Ours 3DGS-MCMC LoD 0 ✓ 25.38 0.716 0.407 4.2 100 421
LoD 1 ✓ 26.19 0.748 0.371 6.1 155 358
LoD 2 ✓ 26.71 0.773 0.339 9.1 242 304
LoD 3 ✓ 27.08 0.791 0.311 13.6 381 256
LoD 4 ✓ 27.32 0.804 0.289 20.6 603 213
LoD 5 ✓ 27.45 0.812 0.273 31.3 962 174
LoD 6 ✓ 27.49 0.815 0.263 46.9 1546 139
LoD 7 ✓ 27.49 0.816 0.259 68.3 2497 117
Ours Scaffold-GS LoD 0 ✓ 24.97 0.710 0.409 4.1 50 144
LoD 1 ✓ 25.98 0.742 0.376 5.5 68 133
LoD 2 ✓ 26.69 0.768 0.345 7.5 94 122
LoD 3 ✓ 27.25 0.789 0.318 10.2 129 108
LoD 4 ✓ 27.62 0.803 0.297 14.0 179 94
LoD 5 ✓ 27.79 0.810 0.284 19.3 248 83
LoD 6 ✓ 27.85 0.812 0.278 26.6 346 74
LoD 7 ✓ 27.87 0.813 0.276 36.9 483 69
Ours Octree-GS LoD 0 ✓ 23.77 0.664 0.447 4.1 50 149
LoD 1 ✓ 24.73 0.697 0.413 5.7 69 140
LoD 2 ✓ 25.61 0.729 0.378 7.8 96 127
LoD 3 ✓ 26.44 0.761 0.341 10.8 133 113
LoD 4 ✓ 27.11 0.789 0.305 14.8 186 100
LoD 5 ✓ 27.55 0.808 0.279 20.5 262 86
LoD 6 ✓ 27.75 0.815 0.265 28.2 371 78
LoD 7 ✓ 27.81 0.817 0.261 38.5 527 73
Context-GS high ✗ 27.57 0.808 0.289 12.4 364 66
med ✗ 27.74 0.812 0.279 18.3 429 58
low ✗ 27.74 0.812 0.279 21.6 477 55
HAC high ✗ 27.17 0.798 0.306 11.5 371 74
med ✗ 27.53 0.807 0.290 15.2 425 72
low ✗ 27.78 0.812 0.277 22.0 501 68
RDO high ✗ 26.45 0.782 0.322 9.7 862 280
med ✗ 26.89 0.796 0.298 15.3 1236 227
low ✗ 27.05 0.801 0.288 23.3 1860 178
Reduced-3DGS high ✗ 27.05 0.807 0.272 22.7 1245 247
med ✗ 27.19 0.810 0.267 29.3 1436 234
low ✗ 27.28 0.813 0.264 45.8 1434 240
Comp-GS high ✗ 26.40 0.778 0.323 9.1 454 233
med ✗ 26.71 0.790 0.305 10.9 470 193
low ✗ 27.28 0.802 0.283 16.4 485 165
SOG high ✗ 26.56 0.791 0.241 16.7 2150 119
low ✗ 27.08 0.799 0.230 40.3 2176 132
EAGLES — ✗ 27.13 0.809 0.278 57.1 1290 156
LightGS — ✗ 27.24 0.810 0.273 51.0 2197 161
Compact-3DGS — ✗ 27.02 0.800 0.287 29.1 1429 108
Method Level Scalable Tanks & Temples
PSNR ↑ SSIM ↑ LPIPS ↓ Size ↓ W ↓ FPS ↑
Ours 3DGS-MCMC LoD 0 ✓ 21.93 0.768 0.334 4.0 100 540
LoD 1 ✓ 22.70 0.792 0.304 5.3 144 453
LoD 2 ✓ 23.22 0.811 0.278 7.2 208 388
LoD 3 ✓ 23.61 0.825 0.256 9.9 301 320
LoD 4 ✓ 23.85 0.835 0.239 13.6 438 260
LoD 5 ✓ 23.99 0.840 0.226 18.7 640 215
LoD 6 ✓ 24.04 0.844 0.218 25.6 938 178
LoD 7 ✓ 24.04 0.844 0.216 34.9 1379 157
Ours Scaffold-GS LoD 0 ✓ 23.13 0.806 0.295 4.1 50 167
LoD 1 ✓ 23.51 0.821 0.273 5.0 61 155
LoD 2 ✓ 23.72 0.832 0.256 6.0 75 145
LoD 3 ✓ 23.86 0.839 0.242 7.4 93 138
LoD 4 ✓ 23.96 0.844 0.231 9.0 114 125
LoD 5 ✓ 24.01 0.848 0.224 11.0 141 116
LoD 6 ✓ 24.04 0.850 0.219 13.5 174 108
LoD 7 ✓ 24.04 0.851 0.218 16.5 215 98
Ours Octree-GS LoD 0 ✓ 22.42 0.776 0.369 4.2 50 147
LoD 1 ✓ 23.14 0.799 0.339 5.7 68 134
LoD 2 ✓ 23.62 0.818 0.310 7.7 93 121
LoD 3 ✓ 23.98 0.834 0.281 10.3 127 107
LoD 4 ✓ 24.25 0.848 0.256 13.7 173 92
LoD 5 ✓ 24.45 0.859 0.235 18.2 235 78
LoD 6 ✓ 24.57 0.866 0.223 24.1 321 66
LoD 7 ✓ 24.61 0.868 0.219 31.0 438 60
Context-GS high ✗ 24.17 0.856 0.215 10.0 222 77
med ✗ 24.30 0.856 0.214 10.0 247 78
low ✗ 24.25 0.855 0.218 10.3 251 77
HAC high ✗ 24.03 0.843 0.237 7.4 261 82
med ✗ 24.05 0.846 0.222 7.9 300 85
low ✗ 24.37 0.853 0.215 11.1 307 88
RDO high ✗ 23.19 0.827 0.253 5.5 395 370
med ✗ 23.16 0.833 0.239 8.0 598 307
low ✗ 23.32 0.839 0.232 11.9 912 257
Reduced-3DGS high ✗ 23.46 0.840 0.228 10.5 558 384
med ✗ 23.55 0.843 0.223 14.0 656 358
low ✗ 23.57 0.844 0.221 20.7 648 360
Comp-GS high ✗ 23.06 0.816 0.278 6.1 270 334
med ✗ 23.32 0.828 0.259 7.3 242 311
low ✗ 23.62 0.837 0.248 9.8 241 286
SOG high ✗ 23.15 0.828 0.198 9.3 1207 216
low ✗ 23.56 0.837 0.186 22.8 1242 227
EAGLES — ✗ 23.20 0.837 0.241 28.9 651 227
LightGS — ✗ 23.55 0.839 0.235 28.5 1211 225
Compact-3DGS — ✗ 23.41 0.836 0.238 20.9 842 147
Method Level Scalable Deep Blending
PSNR ↑ SSIM ↑ LPIPS ↓ Size ↓ W ↓ FPS ↑
Ours 3DGS-MCMC LoD 0 ✓ 29.03 0.886 0.376 4.1 100 652
LoD 1 ✓ 29.39 0.893 0.362 5.5 146 568
LoD 2 ✓ 29.54 0.897 0.351 7.5 212 489
LoD 3 ✓ 29.61 0.900 0.342 10.3 310 419
LoD 4 ✓ 29.64 0.901 0.336 14.1 452 360
LoD 5 ✓ 29.65 0.901 0.331 19.3 660 311
LoD 6 ✓ 29.65 0.902 0.327 26.5 965 263
LoD 7 ✓ 29.61 0.901 0.326 36.1 1411 220
Ours Scaffold-GS LoD 0 ✓ 29.75 0.897 0.358 4.0 50 178
LoD 1 ✓ 30.00 0.901 0.349 4.6 58 172
LoD 2 ✓ 30.15 0.904 0.343 5.4 68 169
LoD 3 ✓ 30.23 0.906 0.338 6.2 80 165
LoD 4 ✓ 30.29 0.908 0.334 7.2 93 161
LoD 5 ✓ 30.33 0.909 0.331 8.4 109 155
LoD 6 ✓ 30.35 0.909 0.329 9.7 127 151
LoD 7 ✓ 30.33 0.909 0.328 11.3 149 148
Ours Octree-GS LoD 0 ✓ 29.38 0.891 0.365 3.7 50 213
LoD 1 ✓ 29.65 0.895 0.359 4.1 56 212
LoD 2 ✓ 29.81 0.897 0.354 4.6 64 212
LoD 3 ✓ 29.91 0.899 0.350 5.1 72 205
LoD 4 ✓ 29.97 0.900 0.347 5.7 81 200
LoD 5 ✓ 30.02 0.901 0.344 6.4 91 195
LoD 6 ✓ 30.03 0.901 0.343 7.1 103 194
LoD 7 ✓ 30.03 0.901 0.342 7.8 116 192
Context-GS high ✗ 30.10 0.907 0.341 3.4 155 121
med ✗ 30.30 0.911 0.332 5.5 167 133
low ✗ 30.27 0.911 0.329 6.7 181 135
HAC high ✗ 29.92 0.903 0.250 3.9 180 166
med ✗ 29.98 0.904 0.340 4.1 187 173
low ✗ 30.34 0.909 0.329 6.3 196 166
RDO high ✗ 29.61 0.903 0.331 7.0 531 335
med ✗ 29.68 0.905 0.323 11.0 891 269
low ✗ 29.72 0.906 0.318 18.0 1478 200
Reduced-3DGS high ✗ 29.63 0.906 0.318 13.6 804 337
med ✗ 29.70 0.907 0.315 18.3 990 304
low ✗ 29.69 0.907 0.314 35.3 988 311
Comp-GS high ✗ 29.22 0.894 0.369 6.0 311 440
med ✗ 29.56 0.899 0.351 7.1 271 315
low ✗ 29.89 0.904 0.336 10.0 246 254
SOG high ✗ 29.12 0.892 0.270 9.3 800 219
low ✗ 29.26 0.894 0.268 17.7 890 236
EAGLES — ✗ 29.75 0.910 0.318 52.4 1192 144
LightGS — ✗ 29.41 0.904 0.329 43.2 956 348
Compact-3DGS — ✗ 29.76 0.905 0.324 23.8 1053 144

Training & Decoding Performance

Encoding and decoding times in seconds; training time in hours. For non-scalable methods, training time refers to a single compression level — obtaining multiple operating points would require training multiple models. GoDe produces all levels within a single training run.

Context-GS HAC RDO-GS Comp-GS
Level Enc ↓ Dec ↓ Train ↓ Enc ↓ Dec ↓ Train ↓ Enc ↓ Dec ↓ Train ↓ Enc ↓ Dec ↓ Train ↓
High 52.7 53.1 1.33 8.6 11.8 1.19 1.1 2.2 1.31 10.8 9.1 1.63
Medium 87.8 82.5 9.7 12.5 2.9 7.4 10.8 9.0
Low 97.2 88.1 14.9 17.9 4.8 19.0 10.9 8.8
Ours 3DGS Ours Scaffold-GS Ours Octree-GS
Level Enc ↓ Dec ↓ Train ↓ Enc ↓ Dec ↓ Train ↓ Enc ↓ Dec ↓ Train ↓
High 10.6 1.2 1.20 2.1 0.2 0.57 5.0 0.2 1.12
Medium 18.1 1.9 same
run		 2.9 0.3 same
run		 6.6 0.3 same
run
Low 32.7 2.4 4.2 0.5 8.8 0.3

Bold Dec values indicate best decoding times. GoDe achieves up to 1762× faster decoding compared to Context-GS at matched compression rates.

Citation

@article{di2025gode, title = {GoDe: Gaussians on Demand for Progressive Level of Detail and Scalable Compression}, author = {Di Sario, Francesco and Renzulli, Riccardo and Grangetto, Marco and Sugimoto, Akihiro and Tartaglione, Enzo}, journal = {arXiv preprint arXiv:2501.13558}, year = {2025} }