Organizations: Department of Statistics, University of South Carolina, Columbia, SC, USA · Department of Statistics and Operations Research, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA · School of Information Technology, Murdoch University, Perth, Australia · Department of Applied Mathematics and Statistics, Johns Hopkins University, Baltimore, MD, USA
Limited training data diversity constrains generative modeling of 3D human bodies: conservative models remain close to observed examples, whereas exploratory models often violate basic body proportions. We introduce a verifier-guided augmentation framework that uses global and mode-local PCA to generate inexpensive candidates, screens them using correspondence-derived skeletal proportions and body-part geometry, and retrains a diffusion model on accepted candidates. Elastic registration provides both the modal structure used by distributed PCA and the dense anatomical correspondence needed for scalable screening without per-candidate body-model fitting. A blinded human study supports the verifier as a conservative gatekeeper, favoring verifier-accepted over rejected outputs. We evaluate full-pool verifier acceptance separately from the coverage and departure of accepted samples and combine them through EAUC. On 4,498 registered DFAUST surfaces, distributed-PCA augmentation achieves 86.32% acceptance, the highest CP-AUC (0.871), and the highest EAUC (0.752), improving EAUC by 32% over real-only and self-augmented diffusion. These results show that mode-local, verifier-guided proposals broaden diffusion generation while maintaining high agreement with calibrated body measurements.
Figures & tables
Figure 1: Two-round verifier-guided augmentation pipeline. GPCA, DPCA, VAE, and real-only diffusion models are first fitted to the real surfaces and evaluated after reference-calibrated screening. Verifier-accepted GPCA, DPCA, and real-only diffusion samples form equal-size augmentation pools; each is combined with the real surfaces to train a new diffusion model from scratch. The VAE remains a first-round comparator.
Figure 2: Mode-based structure of the registered DFAUST population. (a) 2D MDS of pairwise elastic shape distances, colored by the six clusters; gray denotes observations classified as outliers. Circles denote jumping jacks; diamonds denote punching. (b) Modal shape for each cluster.
Figure 3: Representative outcomes from the sequential verifier. From left to right: verifier-accepted; length-flagged; volume-flagged; and arm-area-flagged shapes. Arrows indicate the observed abnormal regions.
Method
N
Accepted
Rate (%)
Global PCA
100,000
13,821
13.82
Distributed PCA
45,000
30,081
66.85
VAE
40,000
24,684
61.71
Diffusion
40,000
37,411
93.53
Diffusion-GPCA
40,000
28,186
70.47
Diffusion-DPCA
40,000
34,527
86.32
Table 1: Overall verifier acceptance for real and generated shapes. N denotes the initial number of shapes; stage-wise rejection counts are reported in Table A1 .
Model
Verifier-Accepted
Verifier-Rejected
Gap
Diffusion
87.9 [83.7, 91.1]
57.2 [51.5, 62.7]
+30.7
Diffusion-GPCA
72.7 [67.4, 77.4]
32.4 [27.4, 37.9]
+40.2
Diffusion-DPCA
84.2 [79.6, 87.9]
59.3 [53.7, 64.7]
+24.8
Diffusion-SelfAug
88.9 [84.9, 92.0]
65.7 [60.1, 70.8]
+23.3
Table 2: Positive human-validity ratings (%) for verifier-accepted and rejected outputs, with descriptive 95% Wilson intervals. Full statistics appear in Table A3 .
Figure 4: Distributional evaluation of 10,000 verifier-accepted shapes per method in the global PCA space. (a) Coverage–precision curves. (b) Coverage and (c) novelty versus radius. Dashed and dotted black curves denote real-split references.
Coverage radii
Novelty radii
Method
r50cov
r70cov
r90cov
r50nov
r70nov
r90nov
Global PCA
6.101
6.753
7.865
7.938
7.242
6.290
Distributed PCA
4.499
5.080
6.524
5.725
5.091
4.174
VAE
4.392
5.097
6.309
6.221
5.506
4.491
Diffusion
1.250
1.481
2.009
1.412
1.219
0.976
Diffusion-GPCA
1.705
2.129
3.114
6.654
2.198
1.386
Table 3: Coverage and novelty quantile radii in global PCA space; bold marks the column optimum.
Method
Acceptance (%)
CP-AUC
EAUC
Global PCA
13.82
0.835
0.115
Distributed PCA
66.85
0.741
0.495
VAE
61.71
0.815
0.503
Diffusion
93.53
0.608
0.569
Diffusion-GPCA
70.47
0.807
0.569
Diffusion-DPCA
86.32
0.871
0.752
Table 4: Verifier acceptance over full candidate pools and CP-AUC over 10,000 verifier-accepted shapes; EAUC is their product. Bold indicates the highest value in each column.
Figure 5: Examples of highly novel, accepted shapes shown above their nearest real neighbors: (a) Diffusion-DPCA. (b) Real-only Diffusion. Examples are randomly selected from the top 1% most novel samples by nearest-real distance.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1: Additional highly novel verifier-accepted outputs shown above their nearest real neighbors. Four outputs per method are randomly selected from the top 1% by nearest-real distance: (a) GPCA; (b) DPCA; (c) VAE; (d) Diffusion-SelfAug; and (e) Diffusion-GPCA.
Initial number
Flagged by filtering metric
Verifier
Acceptance
Source
of shapes
Length
Volume
Arm-area
accepted
rate (%)
Base generators
Global PCA
100,000
80,737
5,204
238
13,821
13.82
Distributed PCA
45,000
11,648
2,664
607
30,081
66.85
VAE
40,000
13,304
1,768
244
24,684
61.71
Diffusion
40,000
1,763
697
129
37,411
93.53
Appendix
Table A1: Detailed sequential verifier outcomes for real and generated shapes. Length-, volume-, and arm-area-specific counts report shapes newly rejected after passing all preceding filtering stages.
Split (%)
Target
Pseudo-generated
CP-AUC
95/5
4,273
225
0.122
90/10
4,048
450
0.214
80/20
3,598
900
0.318
70/30
3,148
1,350
0.394
60/40
2,698
1,800
0.461
50/50
2,249
2,249
0.506
Appendix
Table A2: Complete CP-AUC sensitivity analysis under target/pseudo-generated partitions of the 4,498 real surfaces. Split percentages are reported as target/pseudo-generated.
Verifier-accepted
Verifier-rejected
Setting
Valid (%)
Wilson
Bootstrap
Valid (%)
Wilson
Bootstrap
Gap
Real-only diffusion
87.9 (262/298)
[83.7, 91.1]
[79.0, 94.9]
57.2 (171/299)
[51.5, 62.7]
[53.7, 60.3]
+30.7
Diffusion-GPCA
72.7 (218/300)
[67.4, 77.4]
[62.7, 80.7]
32.4 (97/299)
[27.4, 37.9]
[27.4, 37.2]
+40.2
Diffusion-DPCA
84.2 (250/297)
[79.6, 87.9]
[78.5, 89.6]
59.3 (178/300)
[53.7, 64.7]
[53.3, 65.0]
+24.8
Diffusion-SelfAug
88.9 (265/298)
[84.9, 92.0]
[84.2, 94.0]
65.7 (197/300)
[60.1, 70.8]
[58.0, 71.0]
+23.3
Overall
83.4 (995/1193)
[81.2, 85.4]
[76.4, 89.5]
53.7 (643/1198)
[50.8, 56.5]
[48.5, 57.8]
+29.7
Appendix
Table A3: Full human-evaluation statistics: share of judgments rating shapes valid human bodies, by generation setting and verifier decision, with descriptive 95% Wilson intervals and 95% percentile intervals from a rater-cluster bootstrap (raters resampled with replacement, B=10,000 ). Gap = accepted minus rejected in percentage points.
The Podium Institute for Sports Medicine and Technology University of Oxford Oxford, UK · Visual Geometry Group University of Oxford Oxford, UK · Ukrainian Catholic University Lviv, Ukraine