SWAGE-3D: Spectral Wasserstein Attention Generative Ensemble, A Comparative Analysis on the ShapeNet Dataset

This study proposes SWAGE-3D (Spectral Wasserstein Attention Generative Ensemble), an enhanced 3D-VAE-GAN framework for single-view 3D object reconstruction using voxel-based representations. The proposed model integrates RGB-D encoding, Wasserstein adversarial learning with hybrid Lipschitz regularization, and a self-attention–augmented generator to improve structural coherence and training stability. By combining variational latent modeling with stabilized Wasserstein optimization, the framework aims to address common challenges in 3D generative modeling, including mode collapse, unstable convergence, and insufficient global consistency. The encoder employs a depth-aware feature extraction strategy, while the discriminator utilizes a hybrid spectral normalization and gradient penalty mechanism to ensure robust approximation of the Wasserstein objective. Additionally, an ensemble strategy is applied at inference time to enhance reconstruction reliability. The proposed approach is evaluated on the ShapeNet dataset across 13 object categories using the Intersection over Union (IoU) metric. Experimental results demonstrate a 16.7% improvement over the baseline 3D-VAE-GAN and competitive performance against state-of-the-art voxel-based reconstruction methods. These findings confirm that the synergistic integration of depth cues, stabilized Wasserstein training, and attention mechanisms significantly enhances single-view 3D reconstruction performance.