The rise of hate speech (HS) on social media platforms has become a critical societal issue, often resulting in severe psychological and emotional harm to individuals and communities. Due to multimodal content like images and memes becoming increasingly common, traditional text-based hate speech detection methods encounter significant limitations on social media platforms. This situation allows hate speech to be embedded in visual content, complicating detection for text-only models. Furthermore, existing deep learning-based hate speech detection models addressing multimodal content face a major bottleneck due to limitations in available datasets. This study aims to address these challenges by enhancing hate speech detection in social media images through the utilization of synthetic datasets. The study explored synthetic data generation techniques to create and augment hate speech image datasets. Five models, namely, Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Bidirectional LSTM (BiLSTM) networks, Recurrent Neural Networks (RNNs), and Deep Neural Networks (DNNs), were trained and evaluated using these datasets. The findings reveal that models trained on mixed datasets significantly outperform those trained solely on synthetic or original datasets, achieving a notable balance between precision and recall. For instance, the CNN model demonstrated a substantial accuracy improvement, from 79.17% on the original dataset to 88.18% on the mixed dataset, corresponding to an increase in recall and F1 score. The results obtained underscore the potential of synthetic data as an effective augmentation strategy for generalization across various hate speech scenarios. Moreover, models trained exclusively in synthetic data demonstrated limitations in capturing real-world nuances, highlighting the importance of combining synthetic and original datasets for comprehensive training.