Authors - Nguyen Hoang Kha, Su Truong Phuc, Huynh Tu Anh, Ha Cao Vi, Tran Quoc Hao, Vinh Dinh Nguyen Abstract - Face recognition tasks often suffer from poor performance when limited real training data is available per class. To address this, we propose a novel pipeline that combines LoRA-based fine-tuning of a generative model with synthetic image generation and YOLOv8 classifier training. Our method fine-tunes Stable Diffusion using just a few real face images per identity, then generates realistic, identity-preserving synthetic images to augment the training dataset. The augmented dataset is then used to train a YOLOv8n classifier for face recognition. We evaluate our pipeline on a 31-class face dataset and achieve a Top-1 classification accuracy of 80.05%, significantly outperforming the baseline YOLOv8n model trained only on resized real images (60.94%). These results demonstrate the effectiveness of our method in low-resource settings for face recognition.
