Variational autoencoders have emerged as a powerful tool for representation learning of complex datasets and generative modeling. Despite this, they present limitations in balancing reconstructive performance with representation learning, often learning a poor approximation of the posterior distribution. Various strategies have been employed to combat this, largely focusing on the inclusion of hyperparameters to adjust the balance of the objective functions in favor of representation learning. Recent research has suggested that an evolutionary approach to hyperparameter tuning and objective function balancing may rectify issues that previous methods have failed to solve. This paper presents a new method, named eInfoVAE, which combines a modified objective function with evolutionary tuning to tackle the issue from both perspectives. Experiments on a simple dataset validate that the eInfoVAE is able to outperform previous models, improving reconstructive performance, representation learning, and image generation quality, and able to produce sharp images. eInfoVAE also demonstrates the capabilities of evolution strategies for unsupervised hyperparameter tuning, significantly reducing the time required to tune a model’s hyperparameters.