Abstract Deep learning-based image classification systems necessitate large and balanced datasets for robust generalization, yet real-world scarcity (few-shot) and severe class imbalance critically limit performance, particularly regarding minority classes. This study proposes a ‘Quality-Aware Generative Data Augmentation’ framework integrated with a teacher-based synthetic data filtering mechanism to address these limitations. Utilizing Projected GAN architecture, synthetic images were generated at varying convergence stages and categorized into three distinct quality levels (Sets A, B, and C) extracted from distinct temporal training snapshots (early, intermediate, and convergence stages), yielding empirical FID scores approximately 35, 25, and 10, respectively. To ensure class consistency and semantic reliability, these samples were filtered by a teacher network (ResNet50) fine-tuned exclusively on real data, utilizing a Top-N Selection (Ranking) strategy based on confidence scores. The framework was comprehensively evaluated using MobileNetV3-Large, EfficientNet-B0, and ShuffleNet architectures on two structurally contrasting datasets: the Forest Species Database (FSD) and DermaMNIST. In the FSD dataset, the EfficientNet model’s F1-score improved from 0.881 to 0.937, while the lightweight ShuffleNet achieved a significant 0.11 improvement, rising from 0.792 to 0.903. In DermaMNIST, the filtering strategy successfully mitigated false positives in minority classes, elevating the overall F1-score to 0.8338. The findings indicate that the efficacy of generative data augmentation depends on both data quantity and the generator’s convergence stage combined with the rigour of the selection mechanism. Consequently, this study provides a methodological guideline for optimizing synthetic data usage in data-scarce and imbalanced scenarios, offering a robust solution for data-constrained and lightweight deployment environments.