The primary goal of this work is to systematically evaluate the intra-finger variability of synthetic fingerprints (particularly latent prints) generated using a state-of-the-art diffusion model. Specifically, we focus on enhancing the latent style diversity of the generative model by constructing a comprehensive latent style bank curated from seven diverse datasets, which enables the precise synthesis of latent prints with over 40 distinct styles encapsulating different surfaces and processing techniques. We also implement a semi-automated framework to understand the integrity of fingerprint ridges and minutiae in the generated impressions. Our analysis indicates that though the generation process largely preserves the identity, a small number of local inconsistencies (addition and removal of minutiae) are introduced, especially when there are poor quality regions in the reference image. Furthermore, mismatch between the reference image and the chosen style embedding that guides the generation process introduces global inconsistencies in the form of hallucinated ridge patterns. These insights highlight the limitations of existing synthetic fingerprint generators and the need to further improve these models to simultaneously enhance both diversity and identity consistency.
Figures & tables
Figure 1: Example of a synthetic latent print. The left image shows the source exemplar print; the center image displays the reference latent style (white tape developed with black wetwop); and the right image shows the resulting synthetic latent print.
Figure 2: An illustration of core tasks involved in fingerprint synthesis and how they are handled by the GenPrint framework [ 11 ] .
Method
Source Data
Surfaces
Processing Techniques
Latent Style Control
Wyzykowski et al. [ 26 ]
2,288
NA
NA
No
Joshi et al. [ 13 ]
9,000
40
NA
No
GenPrint [ 11 ]
26,326
37
NA
No
Ours
28,357
40
15
Yes
Table 1: Comparison between our proposed synthetic latent print generation with existing synthetic latent fingerprint generators.
Figure 3: Our proposed modification of GenPrint Stage 2 for diverse latent fingerprint generation with fine-grained style control.
Dataset
# Unique Fingers
# Latent Images
Surface
Latent Processing Technique
In GenPrint’s Training Set?
IIITD-SLF [ 21 ]
120
120
Ceramic
Black powder dusting
Yes
IIITD-MUST [ 15 ]
120
16327
14 surfaces
Black powder, Black wetwop, Cyanoacrylate fuming, DFO, Fluorescent powder, Genetian violet stain, Magnetic black powder, Ninhydrin, White powder, White wetwop
Yes
IIITD-MOLF [ 22 ]
1000
4400
Ceramic
Black powder dusting
Yes
MSP Latent [ 27 ]
1866
2030
Crime Scene
Powder, Chemical, Ink
Yes
NIST-SD27 [ 9 ]
258
258
Crime Scene
NA
No
NIST-SD302 [ 8 ]
1019
3452
22 surfaces
1,2-Indanedione, Black adhesive-side powder, Black powder, Cyanoacrylate, White adhesive-side powder
Yes
Table 2: Summary of real latent fingerprint datasets used for constructing the latent style bank.
Figure 4: Examples of real latent print (left) and synthetic latent print (right) with different latent styles. Here, the style embedding from the real latent is used for generating the synthetic latent.
Figure 5: Visual comparison of two real latent prints developed on black electrical tape using technically identical methods. Despite the same chemical basis and surface, differences in visual appearance can be noticed.
Figure 6: Detailed overview of the finger placement simulator that provides ridge pattern guidance.
TMR (%) @ FMR=0.01%
Surface
Technique
Real Pair
Synthetic Pair
Hybrid Pair
Apple iPhone 5s
Black powder
79
93
99
Black Dustbin Bag
Cyanoacrylate fuming
79
95
90
Ceramic
Black powder
75
95
96
Ceramic Plate
Black powder
86
96
96
Ceramic Plate
Magnetic black powder
89
95
97
Table 3: Verification performance for different latent styles (surface+technique combination) based on MegaMatcher. Here, real pair indicates matching real exemplar and latent prints, synthetic pair indicates matching synthetic exemplar and latent prints, and hybrid pair indicates matching real exemplar and synthetic latents.
Figure 7: Genuine and impostor match score distributions for real and synthetic pairs. In brackets is the number of matching pairs.
Figure 8: The average NFIQ2 quality for synthetic vs. real latent prints. The circle size is proportional to the size of the latent style bank for that style. The dashed line represents the case where the average quality values of real and synthetic latent prints are equal.
Figure 9: Scatter plot correlating the average LFIQA scores of synthetic latent against real latent prints from the same styles.
Figure 10: Examples of NFIQ2 quality score distributions for real and synthetic latent prints corresponding to two styles.
Figure 11: Local error analysis. The first column displays the ground-truth minutiae in red overlaid on the real exemplar image from the IBM-HURSLEY dataset. The second column shows the expected minutiae in red overlaid on the ridge pattern guidance that is fed as input to the ControlNet module. These minutiae are expected to be present in the synthetic print generated by ControlNet. In the third column, the markers illustrate the manual annotation results on the synthetic print: green indicates a matched minutia between the exemplar and synthetic prints, orange indicates missing minutia that is removed by the generator, and purple highlights spurious minutiae introduced during generation.
Figure 12: Qualitative comparison of global hallucination across different sensor styles. The red and green boundaries indicate high and low hallucination scores (1−IoU) , respectively. For all the synthetic prints (middle row), the ridge pattern guidance is the same (top row) but the style guidance is different (bottom row).
Local
Input Quality
Error Type
High
Low
Total
Removal (%)
11.05
13.83
12.42
Addition (%)
10.98
31.67
21.45
Table 4: Local errors (addition and removal of minutiae points) introduced by the generation pipeline.
Figure 13: Examples of extrapolation due to enhancement errors.
Figure 14: Example of global ridge pattern hallucination. Green corresponds to the overlapped region between expected Biexp and synthetic print Bigen masks. Blue shows the hallucinated region.
Sensor
Hallucination Error Rate (%)
CrossMatch
30.04±1.025
Futronic
13.00±0.575
DF90
12.30±0.851
GreenBit
7.30±0.299
Morpho
7.19±0.256
SilkID
6.78±0.448
Table 5: Global hallucination error rates for different sensor styles.