This paper studies whether synthetic fingerprints can supplement scarce real prints from very young children. The authors fine-tuned an existing fingerprint generator on just 205 real prints from nine children aged 0.59–4.23 years. From that fine-tuned model they produced 32,000 candidate images and then applied a strict filtering and verification process. After all checks, 1,985 synthetic fingerprints remained for controlled research use, together with provenance and evaluation records.

The team used a conditional StyleGAN2-style generator that had been pre-trained on 20,844 adult prints from 338 people and then fine-tuned with adaptive augmentation to avoid overfitting. During fine-tuning they relearned class labels for 26 joint finger-and-age categories. Candidates were sampled from several training checkpoints, including a 200-kimg checkpoint that matched the real data distribution well but showed higher similarity to the real reference set. Input real prints were high-resolution images resampled to 500 pixels per inch and placed on a 512-pixel canvas before being used for fine-tuning and evaluation.

Each synthetic candidate was tested with standard fingerprint tools. Image quality was measured with NFIQ 2, which rates how usable an image is for matching. Minutiae—small ridge details that fingerprint matchers use—were extracted with NBIS mindtct and matched with NBIS bozorth3. The authors also checked fingerprint pattern classes using a Poincaré-index method and compared textures using a DINOv2 embedding. Candidates were rejected if they were too similar to any real print in the reference set or if they duplicated other synthetic samples. A final symmetric pairwise verification step removed additional near-duplicates.

Concrete trends in the real pediatric data motivated age-aware evaluation. In the source set the median minutiae count rose from 43 in the 0–1.5-year group to 96 in the 3+ group. Median usable print area increased from 131 to 239 mm2, ridge period from 0.38 to 0.43 mm, and NFIQ 2 quality from 12 to 33.5. The youngest age groups still produced retained synthetic samples within the sampling budget, but the oldest original age group yielded only 24 retained prints. A data-derived two-group age label improved agreement between real and synthetic pattern distributions for the younger children.