Zhejiang University
Starting from the same full model, retain 0.1%–100% of its Gaussians as needed, without retraining.
3DGS represents a scene with many small ellipsoids, each with color, shape, and opacity. The budget K is the number of retained Gaussians; percentages are relative to the full model. One scene asset can serve different device capabilities or changing runtime resources.
CLOD and MGS share parameters across budgets and select ordered prefixes. Optimizing each budget independently requires building separate models. The following approaches all adapt parameters as the count changes, but place the cost at different stages.
Train budget-conditioned networks to select Gaussians and predict geometric changes. Once built, the model supports inference at different budgets without retraining for each one.
Main cost: adaptation training. The selection and transformation networks require training; total offline time is longer in our experiments.
Merge nearby Gaussians and recompute their parameters to obtain a target count directly from an existing model, without training.
Main cost: target-model conversion. A budget model that has not yet been prepared requires simplification computation. Table 2 reports this conversion time.
Directly optimize parameters at a few budgets, then connect the adapted states. Decode the requested budget after construction.
Main cost: extra storage. A residual atlas stores adaptation across budgets to support fast switching.
Training, conversion computation, and storage are distinct trade-offs. Reconstruction quality is a separate comparison: see the images and metrics at matching budgets below.
With fewer Gaussians, each remaining primitive has more of the scene to represent. Its position, shape, opacity, and appearance therefore need to adapt to the budget.
Garden · Test view 12
Every image renders the same 24,000 Gaussians; currently using parameters optimized at a 0.8% budget.
We directly optimize parameters at a few budgets, save these states as anchors, and connect them. Intermediate budgets are decoded from neighboring anchors, without training a separate model for every budget.
Reduce the count and optimize the surviving Gaussians. Each smaller set is selected from the preceding model.
Align the same Gaussian identities across anchors. Store their parameter changes relative to the full model in a residual atlas.
Interpolate shared Gaussian parameters between adjacent anchors. Gate the contribution of entering Gaussians through opacity to produce the requested count.
Scroll horizontally to view the full pipeline ↔
With 11 budget anchors (including Full) , the residual atlas stores each anchor's parameter changes relative to the full model.
Approximately 21.4% of the original full-model PLY size.
Includes training, encoding, calibration, and packaging.
95% of requests take no longer than this (P95). RTX 4090, resident anchors; parameter preparation only, excluding rendering.
Averages cover 13 scenes; additional storage excludes the full-model file.
Compare scene reconstructions, mean metrics over 13 scenes, and quality across continuous budgets. Figures and comparison settings follow the paper.
Scroll horizontally to compare methods ↔
Higher PSNR (pixel reconstruction quality, dB) and SSIM (structural similarity) are better; lower LPIPS (perceptual difference) is better. Tables are vector graphics; scroll horizontally on narrow screens.
We use the official CLOD-3DGS and MGS implementations, with MCMC (Markov chain Monte Carlo) densification in this experiment. CLOD fine-tunes sampled prefixes after full-model pretraining; MGS jointly optimizes the full set and sampled prefixes at each step. ParetoGS adapts parameters at a few budgets and connects the resulting states.
This table compares equal fractions of each method's own full model and reports the actual counts. The full-model sizes differ across methods.
Times beside method names are offline means over 13 scenes. ParetoGS takes about 10.3 minutes for full-model and anchor training, encoding, calibration, and packaging; training accounts for about 9.78 minutes.
All three methods start from the same ImprovedGS full model and are compared at the same requested counts. FlexGS-imp learns budget-conditioned selection and geometric transforms; NanoGS-imp converts a model offline for each target budget; ParetoGS decodes sparse adapted anchors. The suffix -imp denotes the shared initialization.
FlexGS-imp uses the official code with 35,000 adaptation steps and training budgets of 1%, 5%, 10%, and 15%. The evaluated 0.1% and 30% budgets lie outside this range. NanoGS-imp also uses the official implementation.
With a shared full model, ParetoGS outperforms FlexGS-imp and NanoGS-imp on PSNR, SSIM, and LPIPS at all five budgets evaluated in Table 2.
FlexGS-imp's approximately 81 minutes include full-model construction and budget adaptation. NanoGS-imp's 48–56 seconds measure offline conversion from Full to one target model. The demo in Section 03 and millisecond timings in Section 04 concern budget switching after construction.
Inspect quality between anchors against a GNS reference optimized independently at each budget. Curves show means over all test views of the selected scene.
The paper covers the algorithm, component ablations, and complete per-scene results. A BibTeX entry is provided below.
@misc{deng2026paretogshighqualitycontinuousbudget3d,
title={ParetoGS: High-Quality Continuous-Budget 3D Gaussian Splatting with Sparse Adapted Anchors},
author={Xiaobin Deng and Changyu Diao and Min Li and Duanqing Xu},
year={2026},
eprint={2610.34235},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.34235},
}