3D Gaussian Splatting (3DGS) enables photorealistic novel- view synthesis with real-time rendering, yet fine-grained di- agnosis of its rendering artifacts remains challenging. These artifacts are typically spatially localized and exhibit diverse degradation patterns, making them difficult to character- ize using global quality metrics and image-level assessment benchmarks, which provide no explicit indication of where they occur or what types of artifacts they represent. To ad- dress this limitation, we introduce GaussFlaw-80K, the first large-scale dataset for fine-grained perceptual diagnosis in 3DGS. Constructed through a model-assisted human-in-the- loop pipeline, GaussFlaw-80K contains 80,049 artifact re- gions across 14,289 rendered images, each annotated with a pixel-level mask and one of eight flaw labels spanning ge- ometric, blending, and photometric degradations. These an- notations turn 3DGS artifact diagnosis into an explicit and measurable task. Building on this dataset, we propose Ge- oCritic, a geometry-guided multi-task model that integrates RGB appearance with depth-derived cues to jointly localize artifacts and classify their flaw types. Beyond evaluation, we further demonstrate the downstream benefits of fine-grained diagnosis through three complementary targeted operations: Gaussian densification, attribute correction, and outlier prun- ing. Experiments show that GeoCritic achieves an AUC-J of 0.9507 and an Acc@1 of 80.08%, while diagnosis-guided re- finement improves all reported global and crop-level metrics across diverse datasets and 3DGS backbones.
Figures & tables
Name
Scene
Number of Vids/Imgs
Number of Ann.
Annotation Format
Flaw Location
Flaw Type
Refinement
GSC-QA ( Yang et al. 2024 )
15
120
120
Score
×
×
×
GS-QA ( Martin et al. 2025 )
8
64
64
Score
×
×
×
NVS-QA ( Zhang et al. 2025 )
13
65
130
Score
×
×
×
3DGS-IEval-15K ( Xing et al. 2025 )
10
15.2K
15.2K
Score
×
×
×
GaussFlaw-80K (Ours)
26
14.2K
80K
Pixel-level Diagnosis
✓
✓ (8 Types)
✓ (Targeted)
Table 1: Comparison of 3DGS quality assessment datasets. Compared to existing benchmarks, GaussFlaw-80K offers advantages in diversity and volume, featuring 26 complex scenes and 80K pixel-level annotations. By advancing the annotation format to fine-grained diagnosis (enabling explicit flaw location and typing), our dataset uniquely supports artifact analysis and targeted refinement.
Figure 1: Data Construction Pipeline of GaussFlaw-80K. (a) rendered image collection across datasets, 3DGS methods, checkpoints, and novel views; (b) the hierarchical taxonomy of eight artifact categories; (c) the human-in-the-loop annotation pipeline; and (d) annotated examples and dataset statistics.
Figure 2: Overview of the proposed fine-grained perceptual diagnosis pipeline. Left: The GeoCritic model, leveraging a dual-stream architecture for RGB-depth feature extraction and multi-level fusion, followed by a multi-task head for fine-grained diagnosis. Right: The downstream targeted rectification operations guided by the selected artifact regions.
Figure 3: Qualitative comparison of artifact location results. Our method yields sharper and more precise localization than deep saliency detectors and VLM-based methods. More qualitative results are included in the supplementary material.
Model
AUC-J ↑
NSS ↑
CC ↑
SIM ↑
KLD ↓
SR
0.4322
-0.1132
-0.0839
0.2575
2.4141
PFT
0.4370
-0.1070
-0.0794
0.2651
2.1804
SWD
0.4493
-0.2018
-0.0753
0.2998
2.0255
SUN
0.4249
-0.1704
-0.1036
0.2827
2.1543
Murray
0.4726
0.0234
-0.0287
0.3043
2.1249
Salicon
0.9419
1.9878
0.6745
0.5305
1.0941
Table 2: Quantitative evaluation of the GeoCritic for artifact location. Unmarked VLMs are zero-shot, and † denotes LoRA fine-tuning. Best and second-best results are shown in bold and underlined, respectively.
Model
Acc@1 ↑
Acc@3 ↑
M-Recall ↑
M-F1 ↑
Qwen3-VL-8B
27.56%
76.32%
0.1775
0.1479
Qwen3-VL-8B †
75.37%
83.46%
0.6596
0.6748
GLM-4.6V-9B
14.39%
50.00%
0.1106
0.0776
GLM-4.6V-9B †
75.11%
82.87%
0.6230
0.6455
Ovis 2.5-9B
16.23%
54.60%
0.1054
0.0872
Ovis 2.5-9B †
67.12%
77.54%
0.5371
0.5407
Table 3: Quantitative evaluation of GeoCritic for fine-grained topological artifact classification. Unmarked VLMs are zero-shot, and † denotes LoRA fine-tuning. Best and second-best results are shown in bold and underlined, respectively.
Model
Mip-NeRF 360
TandT
Global Quality
Region Quality
Global Quality
Region Quality
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
LPIPS c ↓
DISTS c ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DISTS ↓
LPIPS c ↓
DISTS c ↓
Vanilla 3DGS
27.749
0.823
0.176
0.069
0.231
0.148
23.826
0.853
0.169
0.063
0.222
0.137
+ Ours
27.895
0.828
0.166
0.065
0.206
0.134
23.959
0.858
0.159
0.060
0.201
0.128
Scaffold-GS
28.430
0.833
0.170
0.070
0.223
0.144
23.991
0.850
0.174
0.068
0.207
0.132
+ Ours
28.448
0.838
0.158
0.063
0.195
0.126
24.379
0.857
0.160
0.063
0.180
0.117
Table 4: Quantitative evaluation of GaussFlaw-guided rectification across different 3DGS backbones. Metrics are reported on both complete rendered images (Global) and diagnosed-region crops (Region).
Metric
Baseline
Densification
Correction
Pruning
LPIPS c ↓
0.231
0.212
0.225
0.229
DISTS c ↓
0.148
0.138
0.144
0.146
Table 5: Quantitative analysis of individual diagnosis-guided operations on Mip-NeRF 360 with Vanilla 3DGS. Metrics are computed over diagnosed-region crops.
Figure 4: Qualitative comparison of GaussFlaw-guided rectification across different 3DGS backbones.
School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, China · University of Chinese Academy of Sciences, China · Tsinghua Shenzhen International Graduate School, China