Background: Evaluating embryo quality is crucial for the success of in vitro fertilization procedures. Traditional methods, such as the Gardner grading system, rely on subjective human assessment of morphological features, leading to potential inconsistencies and errors. Artificial intelligence-powered grading systems offer a more objective and consistent approach by reducing human biases and enhancing accuracy and reliability. Methods: We evaluated the performance of five convolutional neural network architectures—EfficientNet-B0, InceptionV3, ResNet18, ResNet50, and VGG16— in grading blastocysts into five quality classes using only embryo images, without incorporating clinical or patient data. Transfer learning was applied to adapt pretrained models to our dataset, and data augmentation techniques were employed to improve model generalizability and address class imbalance. Results: EfficientNet-B0 outperformed the other architectures, achieving the highest accuracy, area under the receiver operating characteristic curve, and F1-score across all evaluation metrics. Gradient-weighted Class Activation Mapping was used to interpret the models’ decision-making processes, revealing that the most successful models predominantly focused on the inner cell mass, a critical determinant of embryo quality. Conclusions: Convolutional neural networks, particularly EfficientNet-B0, can significantly enhance the reliability and consistency of embryo grading in in vitro fertilization procedures by providing objective assessments based solely on embryo images. This approach offers a promising alternative to traditional subjective morphological evaluations.