Facial recognition systems (FRS) in forensic contexts require rigorous proficiency testing, but collecting and sharing real facial images faces increasing privacy and ethical constraints. This thesis explores synthetic facial image generation as a privacy-preserving alternative, investigating whether multiple images preserving a consistent synthetic identity can be generated across varied contexts.

A generation pipeline using Stable Diffusion XL with IP-Adapter FaceID was implemented to generate identity-consistent synthetic facial image sets. The approach was evaluated on 100 synthetic identities, each with one reference and five identity-conditioned images. Identity consistency was assessed through embedding-based metrics and human evaluation by forensic practitioners. Results showed mean identity distances of 0.389 ± 0.039 between reference and conditioned images, with verification performance achieving true accept rates of 95.60 %, 83.40 %, and 75.80 % at false accept rates of 0.10 %, 0.01 %, and 0.001 % respectively.

Perceptual realism assessment using FID scores indicated moderate similarity to real facial images, while forensic expert review revealed varying quality and identity preservation across image pairs. The findings demonstrate that embedding-based identity conditioning can generate synthetic facial image sets suitable for automated verification, while highlighting remaining limitations in fine-grained identity preservation and perceptual realism that require consideration for forensic proficiency testing applications.