3D Gaussian Splatting represents a scene as millions of small, semi-transparent colored blobs that a GPU can draw very quickly. To fit a headset's 11.1 ms frame budget (90 frames per second) or a phone's thermal limits, these scenes are routinely pruned, meaning blobs are deleted until the scene renders fast enough. Every pruning method in the literature is judged the same way: render still test views, then report PSNR or SSIM, two scores that measure pixel-level agreement with a photograph. People, however, see these scenes while moving through them, and pruning plausibly produces artifacts that exist only across frames: blobs popping in and out, flicker on thin structures, and shimmering floaters. A single still frame cannot show any of these.
This note proposes two things. The first is a human study at matched frame times that tests whether pruning methods ranked by still-frame metrics are ranked the same way by viewers watching the scene in motion. The second is a pruning score that weights each blob by how visible its removal would be over a camera trajectory, rather than by its pixel contribution to fixed training views. If the ranking gap exists, the study gives compression research a way to evaluate that respects the real-time constraint. If it does not exist, that is also worth knowing, and the pruning method should be dropped.
Core Claim
We hypothesize the following. At a fixed, measured frame time, human preference among pruned splat scenes viewed in motion disagrees with the ranking produced by per-view PSNR and SSIM. A large part of this disagreement comes from temporal instability that no per-view metric can see. If so, a pruning criterion that weights each blob's pixel contribution by a temporal visibility map should beat contribution-only pruning in human preference at the same frame time, even when its PSNR is equal or lower.
What is new is the combination: evaluation at a matched frame-time budget, stimuli viewed in motion, and human judgments on full unbounded scenes. Prior work supplies each piece separately but never together.
Prior Work and Its Role
Three papers supply the methods we compare against. All of them prune or allocate blobs using signals defined on single views.
- RadSplat keeps the blobs that contribute most to pixels along training rays. It reaches 10x fewer blobs and over 900 FPS, and reports SSIM.
- Perceptual-GS steers where blobs are added using Sobel edge maps. It uses 15% fewer blobs and reports PSNR and LPIPS, but no human judgments.
- 3D-HGS speeds up rendering by culling the transparent halves of split kernels.
We reuse RadSplat's contribution score as our starting point and change only how pixels are weighted.
Two papers motivate the proposal. Perceptual Quality Assessment of 3D Gaussian Splatting shows that SSIM correlates with human opinion scores at a Pearson correlation (PLCC) of only 0.2936, against 0.8213 for their learned no-reference metric, GSOQA. That study used 225 static, bounded objects, and downsampling was one of its distortion types. It therefore establishes that divergence exists, but not whether it holds for unbounded scenes in motion under a frame budget.
Our choice of a temporal visibility model comes from Instant Reality. That work schedules mesh streaming with a temporal perceptual model (FovVideoVDP), explicitly because level-of-detail swaps cause popping that users notice. It works on meshes, not splats.
The audio papers constrain what we claim in general terms. ViSQOL v3 and SongEval show that learned metrics fitted to human data can themselves be unreliable proxies; Audiobox-Aesthetic reaches only 0.61–0.66 correlation. This is why we validate every learned metric, GSOQA included, against new human data rather than optimizing against it blindly.
Research Program
Budget matching. Methods are compared at equal measured frame time on fixed hardware, not at equal blob count. The cost of sorting and blending blobs depends on how much they overlap on screen, so blob count is a poor proxy for speed. For each method we binary-search its pruning ratio until the median frame time along a test trajectory hits the target. We use two targets: 11.1 ms on a mobile-class GPU and 5 ms on a desktop GPU.
Stimuli. We render smooth camera paths that interpolate between held-out views on Mip-NeRF 360, Tanks & Temples, and Deep Blending. Each clip lasts a few seconds and is shown both on a monitor and in a headset.
Human study. Participants make paired forced choices ("which clip looks better?") between methods at the same budget. Pairwise preferences are converted to per-method quality scores. This yields a dataset of budget-matched, in-motion judgments for splat pruning, which does not currently exist.
Metric audit. For each metric we compute pairwise agreement with the human majority and Kendall's rank correlation. The metrics are per-view PSNR, SSIM, LPIPS, GSOQA (if its released code runs on scene-scale models), and a temporal video visibility metric of the FovVideoVDP type.
Proposed pruning score. The score for blob $i$ is
$$ s_i = \sum_{t} \sum_{p} w_t(p)\, c_i(t,p). $$
Here $t$ ranges over frames sampled along jittered trajectories near the training views, and $p$ over pixels. $c_i(t,p)$ is blob $i$'s alpha-blending contribution to pixel $p$, the same quantity RadSplat uses. $w_t(p)$ is a per-pixel visibility map from the temporal perceptual model, computed once offline between the full-scene rendering and a coarsely pruned rendering.
With uniform $w$, this score reduces to RadSplat pruning. We prune by $s_i$ and then fine-tune briefly, exactly as the baselines do. All of the cost is offline, so frame time at render is unaffected.
Baselines
Uniform random pruning. The naive reference point is to remove blobs uniformly at random from a vanilla 3DGS model until the frame-time target is met, followed by the same fine-tuning. Any method that fails to beat random pruning in human preference at equal frame time is not doing useful work.
RadSplat. The strongest existing method is RadSplat's contribution pruning, with its Light variant as the aggressive-budget setting. It is the most direct competitor because it already optimizes exactly the quantity our score reweights. Its reported result is SSIM 0.843 on Mip-NeRF 360, above vanilla 3DGS at 0.815, at over 900 FPS on an RTX 3090 (RadSplat). No human-preference result for any splat pruning method under a frame budget was found on Lacuna.
To keep the comparison fair, RadSplat and our method will share:
- the same trained starting model,
- the same trajectory sampling,
- the same fine-tuning iterations,
- the same budget search.
We will reimplement RadSplat's pruning on a shared codebase rather than compare against its reported numbers.
Perceptual-GS. Perceptual-GS is the strongest method that is already perception-motivated. It reports PSNR 28.01 against 27.71 for vanilla 3DGS on Mip-NeRF 360, with 15% fewer blobs (Perceptual-GS). It will be trained as published and then pruned to each budget by opacity.
Ablation of the temporal ingredient. This ablation replaces $w_t$ with a per-frame LPIPS spatial error map. Everything else stays the same, including the pixel-weighting mechanism.
What counts as beating a baseline. Our method beats a baseline if it has a statistically significant human preference over it at both budgets, on a majority of scenes, at matched frame time. It must also beat the LPIPS-weighted ablation; otherwise the temporal component is not what matters.
Why not just…?
Why not just fine-tune the pruned model with an LPIPS loss? This addresses sharpness in individual frames but gives no signal about flicker between frames, because LPIPS never compares consecutive frames. We will include LPIPS fine-tuning applied on top of RadSplat pruning as an extra arm. If this arm matches our method in human preference, then the pruning criterion does not matter and we will report exactly that.
Why not just prune against GSOQA directly? GSOQA is the best human-correlated splat metric available. However, it was fit to 225 static objects (Perceptual Quality Assessment of 3D Gaussian Splatting), and optimizing a small learned metric invites exploiting its blind spots, a failure pattern the audio literature documents for learned aesthetic scorers (SongEval). Our metric audit measures GSOQA's agreement with humans on scene-scale clips viewed in motion. If that agreement is high, a GSOQA-guided pruning arm becomes a serious competitor and will be added.
Why not just use Sobel-based perceptual allocation, as Perceptual-GS does? Edge maps mark where detail exists, not where removing detail is visible in motion. Thin foliage is both high-edge and flicker-prone, so the two signals may coincide there, which makes this a real threat. Perceptual-GS is therefore run as a full baseline, and a Sobel-weighted $w$ is added as a second ablation.
Why not just report more metrics instead of running a user study? The central question is whether the metrics are right about human judgment, so the metrics cannot be the judge.
Decisive Checks
The first check is whether per-view PSNR and SSIM rankings disagree with human rankings at matched frame time in a meaningful fraction of scene-and-budget pairs. If agreement is near perfect, the premise fails and the pruning method is dropped.
The second check is whether the temporal metric predicts human preference better than per-view LPIPS on the same clips.
The third check is whether our score beats both RadSplat and the LPIPS-weighted ablation in human preference at both budgets, including in headset viewing.
A secondary diagnostic tests whether the gap concentrates in high-detail regions such as foliage and gravel, where the 3DGS-QA study found human sensitivity heightened.
References
- RadSplat: Radiance Field-Informed Gaussian Splatting for Robust Real- Time Rendering with 900+ FPS
- Perceptual-GS: Scene-adaptive Perceptual Densification for Gaussian Splatting
- 3D-HGS: 3D Half-Gaussian Splatting
- Perceptual Quality Assessment of 3D Gaussian Splatting: A Subjective Dataset and Prediction Metric
- Instant Reality: Gaze-Contingent Perceptual Optimization for 3D Virtual Reality Streaming
- ViSQOL v3: An Open Source Production Ready Objective Speech and Audio Metric
- SongEval: A Benchmark Dataset for Song Aesthetics Evaluation