On December 12, 2025, Márk Gyöngyösi successfully defended his MSc thesis entitled „Implementation of In Context Learning for Facial Image Quality Assessment Evaluation and Compliance“.
Abstract:
Facial image quality assessment (FIQA) is essential for ensuring reliable performance in face recognition (FR) systems. Existing FIQA algorithms estimate quality through feature vectors but provide limited interpretability. This thesis investigates whether vision capable Large Language Models (LLMs) can complement existing FIQA workflows. A set of state of the art LLMs were evaluated on a dataset containing face images. Models were tested on four tasks: (1) binary classification of standard compliance, (2) identification of specific defects, (3) prediction of quality scores comparable to FIQA generated quality scores, and (4) generation of naturallanguage explanations that justify the model’s decision. Multiple prompt variants were tested to measure the effect of instruction detail on performance. The results show that larger models benefit from extended context, whereas smaller models frequently misinterpret instructions or deviate from the expected result. Across models, subjectrelated defects (e.g., face obstructions, expression) are detected more consistently than capturerelated defects (e.g., illumination, resolution, background). Qualitative analysis further reveals that explanations remain fluent and confident even when incorrect. Overall, the study demonstrates that while LLMs can produce coherent and interpretable feedback, their reliability is inconsistent and strongly dependent on prompt design and model scale. Nonetheless, their ability to provide accurate assessments—despite not being trained specifically for FIQA, highlights the remarkable generalization and in context learning capabilities of modern LLMs. Although they cannot yet replace human experts or dedicated FIQA systems, these results suggest that with targeted training or task specific fine-tuning, multimodal LLMs hold considerable potential for future FIQA applications.