<p>This paper presents a comparative study of three major families of deep generative models—WGAN-GP, β-VAE, and diffusion models (DDPM)—applied to 3D shape generation. Using the standardized ModelNet40 dataset, we evaluated the architectures along four main axes: geometric fidelity (measured via FID3D), morphological diversity, computational efficiency (training and inference time, memory), and stability of convergence. The results show that diffusion models consistently achieve the best geometric fidelity, albeit with the highest computational cost. β-VAE provides the fastest training and inference with high diversity, but at the expense of reduced fidelity. WGAN-GP offers an intermediate compromise between fidelity, diversity, and computational load, representing a pragmatic option in resource-constrained scenarios. Ablation experiments conducted on DDPM highlight the influence of latent size, diffusion steps, β regularization, and voxel resolution on performance, revealing clear trade-offs between fidelity gains and computational overhead. Overall, the study confirms that no single architecture dominates across all criteria, and suggests that the optimal choice of model depends on the target application and the balance required between fidelity, diversity, and efficiency.</p> Graphical abstract <p>AI-Driven 3D Shape Generation for Additive Manufacturing</p>