This research addresses whether AI models trained with AI-enhanced low-detail images can lead to better performance compared to models trained exclusively on conventional but low-detailed data. Initially, the low-detail image is processed by a pre-trained GAN model (GFPGAN), specially trained to restore low-detailed facial images to higher-quality ones. The dataset SFEW is used as a base, and all its images are enhanced through GFPGAN to make a separate enhanced dataset. The two datasets are further resized to \(48 \times 48\) , \(96 \times 96\) , \(150 \times 150\) , \(250 \times 250\) . Subsequently, a Convolutional Neural Network (CNN) is trained separately on every resolution’s enhanced and normal-quality images. Along with the SFEW test datasets, the \(48 \times 48\) model is also tested on test data from the FER-2013 dataset. The result assesses the performance of each model across seven basic facial expressions, i.e., fear, disgust, anger, surprise, sadness, happiness, and neutral, as well as the overall performance of the model against the whole dataset. The results reveal better accuracy of the models trained on enhanced images than those trained on normal images.