Benchmarking MLLM Robustness in Face Verification: A Comparative Study of Preprocessing and Image Enhancement Effects

The evolution of Multimodal Large Language Models (MLLM) offers new visual capabilities, but their reliability in facial biometrics has not yet been thoroughly researched. This study evaluates the robustness of ChatGPT-4o, Gemini 2.5 Flash, and Qwen3 VL on the LFW dataset through three stages of experimentation: baseline, facial cropping variations, and enhancement techniques. Results show varying outcomes across models, with ChatGPT-4o and Gemini superior on original images that preserve context, while Qwen3 VL requires specific cropping (YOLOv8) for optimal results. A crucial finding shows that sharpening techniques consistently improve accuracy across all models, with ChatGPT-4o achieving a peak performance of 99.20% accuracy with sharpening enhancement and 8 wrong predictions. In contrast, the MediaPipe algorithm was found to significantly degrade performance using Gemini 2.5 with 75.30% accuracy and 247 wrong predictions. This study concludes that unlike conventional models, MLLM prioritizes texture sharpness and macro context over strict facial alignment.