A deep generative approach to personalized super mario level design.

Designing game levels that appropriately match a player's skill level remains a fundamental challenge in procedural content generation, as mismatches in difficulty can lead to player boredom, frustration, and reduced engagement. While deep generative models enable automatic level synthesis at scale, the comparative effectiveness of different GAN architectures for skill-conditioned and personalized level generation remains insufficiently explored. In this work, we investigate the use of generative adversarial networks (GANs) for skill-conditioned procedural content generation by evaluating five architectures: U-Net GAN, StyleGAN, Deep Convolutional GAN (DCGAN), ResNet-GAN, and Spectral Normalization GAN (SN-GAN). Player behavior is clustered into discrete skill groups using Spectral Clustering, and the resulting labels are employed as conditioning signals for level generation. The selected architectures span different convolutional depths, normalization strategies, and regularization mechanisms, enabling a broad and systematic comparison. All models are evaluated using consistent quantitative metrics, including tile distribution entropy, diversity score, discriminator accuracy, generation speed, and pairwise Hamming distance. Quantitative analysis demonstrates that as player skill increases, generated levels exhibit fewer deaths and shorter completion times, while encouraging more complex interactions such as higher jumps and coin collection. Experimental results indicate that ResNet-based and U-Net-style GAN architectures achieve the most favorable balance between level diversity, playability, and training stability, corroborating both qualitative and quantitative assessments. The spectral clustering-based skill-conditioning approach successfully produced personalized levels aligned with each player's abilities, supporting adaptive procedural content generation across different skill groups.