Evaluation of latent diffusion enhanced face recognition under forensic image degradations

Abstract Forensic face recognition systems face critical performance degradation when processing low-quality evidence imagery. This study presents a systematic evaluation of latent diffusion-based image enhancement as a preprocessing step for face recognition across seven forensically relevant degradation categories. Using 3,000 individuals from the Labelled Faces in the Wild dataset, we conduct 48,000 recognition attempts in total across eight conditions and compare our pipeline against ESRGAN, Real-ESRGAN, and CodeFormer baselines. Our enhancement pipeline is based on Flux.1 Kontext Dev with Facezoom LoRA adaptation, classifier-free guidance, and optimised DDIM sampling, raises overall recognition accuracy from 29.1% to 84.5%, a 55.4% point improvement with a 95% confidence interval of 51.8 to 59.0 percentage points and a Cohen d value of 2.31. Biometric evaluation confirms these gains are not artefacts of threshold selection: the area under the ROC curve improves from 0.641 to 0.951, the equal error rate falls from 38.2% to 9.1%, and the true accept rate at a false accept rate of 1 in 1,000 rises from 0.187 to 0.731. Per-degradation analysis reveals exceptional recovery for Gaussian blur at 85.7 percentage points and more constrained improvement for JPEG compression at 47.6 percentage points. Cross-architecture evaluation with FaceNet512 confirms consistent directional improvement across both backbones, with AUC gains of 0.271 for ArcFace and 0.283 for FaceNet512, establishing that results reflect genuine image quality enhancement rather than model-specific behaviour. The proposed pipeline achieves 14.7 percentage points higher accuracy than CodeFormer, providing quantitative evidence for diffusion-based enhancement in forensic face recognition.