Voxel representations have long been adopted for structured 3D modeling due to their regular grid formulation and compatibility with convolutional neural networks. Early works explored visual similarity and rotation-invariant normalization for 3D retrieval [26,27].

The emergence of deep generative modeling significantly advanced volumetric synthesis. Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) established probabilistic and adversarial paradigms for high-dimensional data generation. Building upon these foundations, Wu et al. introduced 3D-GAN and 3D-VAE-GAN [28], enabling volumetric generation from both latent codes and 2D images. Recent advancements have sought to unify these paradigms; for instance, UniRecGen [29] integrates feed-forward reconstruction with diffusion-based generation to achieve highly consistent multi-view 3D models.

Recurrent aggregation mechanisms were later proposed for multi-view reconstruction in 3D-R2N2 [30]. Improvements in adversarial stability and scalable generative modeling were introduced through Wasserstein formulations like 3D-IWGAN [31] and geometry-aware 3D GANs [32]. Additional voxel-based generative and classification studies provided further architectural insights into 3D representation learning and recognition [33–37]. In particular, orientation-aware modeling strategies also enriched the design space of 3D deep learning frameworks [38].

More recent voxel-based frameworks focus on improved feature fusion and long-range dependency modeling. Pix2Vox [39] introduced context-aware fusion for combining coarse reconstructions, while TMVNet [40] leveraged Transformer-based encoders for enhanced multi-view feature aggregation. Similarly, VoxFormer [41] demonstrated the power of sparse voxel transformers for camera-based volumetric scene completion. To further refine these architectures, SS3DNet-AF [42] proposed an attention-based fusion mechanism that enhances the network’s focus on relevant geometric details from a single view. Similarly, recent efforts have directly upgraded multi-view pipelines by integrating multi-head attention refiners to reduce boundary prediction errors and capture intricate structural nuances [43]. Attention-enhanced voxel models such as SV3D-CDFF [44] and Semantic Voxel Structure (SVS) [45] further addressed discrepancies between image and voxel domains.

To overcome voxel resolution limitations, alternative 3D representations have been explored. Point-cloud-based generative models employed Earth Mover’s Distance and deep autoencoding strategies for shape generation and completion [46,47]. DescriptorNet introduced an energy-based probabilistic model for 3D pattern synthesis [48].

Implicit representations further advanced continuous shape modeling. Occupancy Networks [49] represented 3D objects as continuous decision functions, enabling high-resolution reconstruction without explicit voxel discretization. IF-NET [50] extended implicit learning for shape completion. While these methods, alongside recent accelerated rendering advancements [51,52], alleviate discretization constraints, voxel representations remain advantageous for direct convolutional processing and adversarial volumetric learning.

Beyond implicit and point-based paradigms, studies have explored deformable surface modeling and graph-based generative approaches for structured 3D synthesis. Multi-chart surface parameterization and surface-oriented generation methods demonstrated that complex 3D geometry can be reconstructed through mesh- or chart-based learning strategies [53–55]. Studies on dynamic reconstruction, semantic shape modeling, and volumetric autoencoding supported the feasibility of learning structured geometric representations under diverse architectural assumptions [56–59]. While these approaches offer advantages in continuous surface modeling, they often require predefined topology constraints or specialized deformation priors. In contrast, voxel-based representations provide a regular volumetric grid structure that facilitates stable adversarial training and unified convolutional processing within a single-view reconstruction pipeline.

Recent advances in point-based and geometric learning have further expanded 3D representation capabilities. Transformer-based architectures and Vision Transformer (ViT) adaptations have been applied to point cloud processing and large-scale generalizable reconstruction to model global spatial dependencies more effectively [60,61]. In parallel, geometric neural operator frameworks have emerged as powerful tools for learning structured mappings in high-dimensional geometric domains [62,63]. These developments provide promising alternatives for geometric representation learning, complementing voxel-based approaches in scenarios where continuous or sparse representations are preferred. Transformer-based 3D reconstruction has also been extended toward neural implicit surface modeling. For example, SparseNeuS [64] integrates transformer architectures with neural implicit surfaces to enhance geometric consistency across sparse views, demonstrating improved structural coherence in continuous representations.

Depth information has increasingly been incorporated to enhance geometric inference. Multi-view supervision methods such as Differentiable Ray Consistency (DRC) [65] demonstrated that volumetric predictions can be learned without explicit 3D labels. Recent advances in monocular depth estimation, particularly Depth Anything V2 [66], have significantly improved robustness and generalization across diverse environments.

Architectural refinements have also focused on stability and feature modeling. Self-attention mechanisms and Transformer-based encoders have been introduced to better capture long-range spatial dependencies in 3D reconstruction [67–69]. Ensemble learning strategies have been widely adopted to reduce prediction variance and enhance robustness in deep neural networks.

Despite substantial progress, several challenges persist in single-view voxel reconstruction, including limited geometric cues in RGB-only inputs, adversarial instability in high-dimensional volumetric generation, and reconstruction variance across training epochs. Although prior studies have addressed these issues individually, their combined and systematically stabilized integration within a unified RGB-D voxel-based generative framework has received comparatively limited investigation. In this context, the present study emphasizes the controlled integration and empirical validation of complementary stabilization strategies for depth-assisted voxel reconstruction.

Unlike our previous VAE-based approach [70], which primarily focused on transfer learning for latent representation stabilization, the proposed SWAGE-3D framework systematically integrates adversarial Wasserstein training, spectral normalization, self-attention modules, and ensemble learning to address stability and geometric coherence in single-view voxel reconstruction. SWAGE-3D framework builds upon the foundational work of Wu et al. [28] on 3D-VAE-GAN, introducing several key innovations to enhance both training stability and the quality of generated 3D models. The overall architecture, illustrated in Fig. 2, consists of three primary network components that operate in a unified voxel-based generative pipeline.

The architecture integrates three primary network components that work together to deliver superior performance. The Encoder Network serves as the initial processing unit, taking 2D input images that include both RGB and depth information and mapping them into a latent space characterized by mean (μ) and variance (σ) parameters. By incorporating depth maps alongside RGB data, the encoder captures richer geometric details, enabling more accurate and informative latent representations. The Generator Network then transforms latent vectors sampled from this space into detailed 3D voxel models using transposed 3D convolutions and a novel self-attention mechanism. This mechanism enhances the model’s ability to capture long-range structural coherence. Finally, the Discriminator Network evaluates the generated 3D models to differentiate between real scanned objects and synthetic samples. By providing precise adversarial feedback based on the Wasserstein distance, this component guides the generator in producing outputs that are increasingly indistinguishable from real-world data. The algorithm for the proposed model is shown in Algorithm 1.

The SWAGE-3D framework employs a depth-assisted RGB-D encoding strategy to enrich geometric feature representation from single-view inputs. Instead of relying solely on RGB images, monocular depth maps are generated and fused as an additional channel, forming a four-channel input representation. Depth maps are extracted from the input RGB images using the Depth Anything V2 model, which provides robust monocular depth estimation across diverse scenes. Each RGB image of size 224×224×3 is augmented with its corresponding depth map to construct a 224×224×4 RGB-D tensor. This integration enables the encoder to capture complementary geometric cues that are not explicitly encoded in color information alone.

All inputs are normalized using ImageNet statistics for the RGB channels, while the depth channel is normalized using its dataset-wide mean and standard deviation to maintain scale consistency. A ResNet18 backbone is adopted as the encoder component. Since the standard ResNet18 architecture is pre-trained on3-channel RGB images, its first convolutional layer is modified to accept 4-channel RGB-D inputs. To preserve transfer learning benefits, the pre-trained weights corresponding to the three RGB channels are retained. The weights of the additional depth channel are initialized as the mean of the RGB channel weights: Wdepth=13(WR+WG+WB). This initialization allows the depth channel to start from a balanced representation without disrupting learned feature distributions. The modified encoder outputs mean μ and log-variance log⁡σ2 parameters, from which latent vectors are sampled via the reparameterization trick:(1)z=μϕ(x)+σϕ(x)⊙ε,ε∼𝒩(0,I)where ε stands for the random noise drawn from the standard normal distribution. This stochastic encoding enables variational learning while preserving stable gradient propagation during training.

The generative backbone consists of a 3D convolutional generator and a Wasserstein-based discriminator. The generator transforms the sampled latent vector into a 32×32×32 voxel grid. To address instability commonly observed in standard GAN training, a Wasserstein GAN with Gradient Penalty (WGAN-GP) formulation is adopted [71]. The discriminator estimates the Wasserstein distance between real and generated voxel distributions. Its objective is defined as:(2)ℒD=Ex^∼Pg[D(x^)]−Ex∼Pr[D(x)]+λGPEx~∼Px~(‖∇x~D(x~)‖2−1)2

Although the Wasserstein distance formulation based on the Kantorovich–Rubinstein (KR) duality provides a theoretically sound objective, its practical implementation remains an approximation [72]. In theory, the KR dual formulation considers the supremum over all 1-Lipschitz functions. However, the critic is restricted to a parameterized neural network with finite capacity. The learned critic provides only an approximation of the true Wasserstein-1 distance rather than its exact computation [72].

To mitigate these limitations, we adopt a hybrid stabilization strategy combining spectral normalization (global Lipschitz control) with a reduced-weight gradient penalty term (local gradient regularization). This dual constraint balances critic expressiveness and stability, improving approximation smoothness without severely restricting model capacity. This approach serves as a stabilized practical realization of Wasserstein-based adversarial learning. In our generator network, a self-attention mechanism is incorporated after the third layer to model long-range spatial dependencies, ensuring global consistency in the generated volumes.

To enhance structural coherence in volumetric generation, a self-attention mechanism is incorporated into the generator. While convolutional layers effectively capture local spatial patterns, their receptive fields remain limited. In voxel-based reconstruction tasks, long-range spatial dependencies are critical for maintaining global object consistency. The self-attention module enables direct interactions between distant spatial regions within the volumetric feature maps. For an intermediate 3D feature tensor F∈RC×H×W×D, query, key, and value representations are obtained via 1×1×1 convolutions. The attention map is computed based on query-key similarity, and the resulting representation is combined with the original feature tensor through a residual connection: F′=γ⋅Attention(Q,K,V)+F. Here, γ is a learnable scalar parameter initialized to zero to stabilize early training.

To enhance inference robustness and mitigate performance fluctuations across training epochs, a checkpoint-based ensemble strategy is employed during the testing phase. Instead of relying on a single trained model, multiple late-stage checkpoints are aggregated to produce a more stable and accurate voxel reconstruction. Model performance was monitored using the IoU metric, and a stabilization plateau was observed between epochs 125 and 149, indicating convergence with minor oscillations. From this stable region, epochs 129, 139, and 149 were selected to construct the ensemble.

Let Gi(x) denote the voxel prediction produced by the i-th checkpoint model, and let wi represent its validation IoU score. The final ensemble prediction is computed as a weighted aggregation: V^(x)=∑wiGi(x)∑wi. Following aggregation, a voxel occupancy threshold τ=0.5 is applied to obtain the final binary volumetric reconstruction. This ensemble mechanism reduces variance introduced by stochastic training dynamics and improves overall reconstruction accuracy.

The optimization of SWAGE-3D is formulated as a multi-stage composite objective that jointly regulates latent space regularity, volumetric reconstruction fidelity, and adversarial distribution alignment. The framework simultaneously optimizes three interacting networks: the encoder (E), the generator (G), and the discriminator (D), using a coordinated update strategy that bridges Variational Autoencoder (VAE) and Generative Adversarial Network (GAN) paradigms.

The primary objective of the VAE stream is to ensure that the encoder-generator pair can effectively map a single-view RGB-D input to its corresponding 3D geometry. To ensure voxel-wise consistency between predicted grids and ground-truth occupancy, a Mean Squared Error (MSE) reconstruction loss is employed:(3)ℒrec=1N∑i=1N(Vipred−Vigt)2where Vpred and Vgt denote the predicted and ground-truth voxel grids, respectively. Simultaneously, the encoder is regularized using the Kullback–Leibler (KL) divergence to enforce the latent distribution to follow a unit Gaussian prior p(z)∼𝒩(0,I) [73]:(4)ℒKL=−12∑(1+log⁡σ2−μ2−σ2)

This dual objective allows the encoder to capture a robust, continuous representation of the 3D shapes while the generator learns the fundamental volumetric mapping.

To enhance the realism and structural sharpness of the generated voxels beyond simple point-wise matching, an adversarial stream is integrated. Under the Wasserstein formulation, the discriminator acts as a critic that estimates the Earth-Mover distance between the generated distribution Pg and the real data distribution Pr. Its objective is defined as:(5)ℒD=Ex^∼Pg[D(x^)]−Ex∼Pr[D(x)]+λGPℒGPwhere ℒGP represents the gradient penalty term enforcing the 1-Lipschitz continuity constraint [74]. In our implementation, spectral normalization is applied to the critic’s layers to provide global Lipschitz control, allowing the gradient penalty coefficient λGP to be reduced to 1.0. This prevents the vanishing gradient problem and stabilizes the training of the generator.

The technical novelty of the SWAGE-3D training procedure lies in the coordinated update of the generator. The generator is not merely trained to fool the discriminator but is explicitly forced to satisfy both the VAE reconstruction constraint and the GAN adversarial requirement. The total objectives for the encoder (ℒE) and generator (ℒG) are defined as:(6)ℒE=λrecℒrec+λKLℒKL (7)ℒG=λrecℒrec−Ex^∼Pg[D(x^)]

By minimizing ℒG, the generator learns to produce voxels that are geometrically accurate relative to the input image (via ℒrec) while simultaneously conforming to the global distribution of real 3D objects (via the adversarial term). In all experiments, the weights are set to λrec=1.0 and λKL=1.0. To maintain equilibrium in this complex adversarial landscape, the discriminator is updated ncritic=5 times for every single update of the encoder and generator, ensuring a reliable approximation of the Wasserstein distance before each generative refinement step.

To evaluate the robustness of the SWAGE-3D framework, we conducted a sensitivity study on the primary loss weights, λrec and λKL, using the Loudspeaker category as a representative sample. We tested scaling factors of ±2× relative to the default configuration. Due to computational constraints, each configuration was trained for 30 epochs to observe early convergence trends and training stability. The results are summarized in Table 2.

The results demonstrate that SWAGE-3D exhibits remarkable stability across different objective weightings, with IoU scores remaining within a narrow margin (<2%). While the ‘High Reconstruction’ (λrec=2.0) configuration yielded a slightly higher IoU of 0.6766 at this stage, it introduced minor oscillations in the discriminator loss, which could potentially destabilize the adversarial balance during full-scale 150-epoch training. Conversely, reducing λKL to 0.5 led to increased variance in the Wasserstein critic’s feedback, confirming that adequate latent regularization is essential for smooth adversarial mapping. Doubling the KL weight (λKL=2.0) maintained stability but resulted in slightly smoother, less detailed voxel grids (posterior collapse). Based on these observations, the λrec=1.0 and λKL=1.0 configuration was selected as the optimal balance for long-term training, ensuring both high geometric fidelity and consistent adversarial equilibrium across all object categories.

To quantitatively evaluate the reconstruction performance of SWAGE-3D and the compared methods, multiple complementary metrics were employed to assess volumetric overlap, structural accuracy, and geometric distance consistency. The primary evaluation metric is the Intersection over Union (IoU), which measures the volumetric overlap between predicted and ground-truth voxel grids. IoU is defined as:(8)IoU=∑i(Vipred⋅Vigt)∑i(Vipred+Vigt−Vipred⋅Vigt)where Vpred and Vgt denote the predicted and ground-truth voxel occupancy grids, respectively. This metric provides a rigorous assessment of spatial consistency between reconstructed and reference 3D shapes. For all evaluations, a voxel occupancy threshold of τ=0.5 was applied to binarize the generator outputs.

While IoU evaluates total volumetric overlap, it may not fully capture the precision of reconstructed surface details. Therefore, the F-score (F1) is additionally reported to provide a balanced measure of reconstruction fidelity by calculating the harmonic mean of precision and recall based on voxel occupancy. The F-score is computed as:(9)F1=2⋅Precision⋅RecallPrecision+Recall

In this context, precision represents the accuracy of the predicted occupied voxels, while recall indicates the fraction of the ground-truth structure successfully recovered by the model. This metric is particularly effective in evaluating the reconstruction quality of thin and complex structures where volumetric overlap alone might be insufficient to reflect structural integrity.

To further quantify the geometric consistency between reconstructed and ground-truth shapes, the Chamfer Distance (CD) is employed. Given two point sets P and G derived from the coordinates of occupied voxels in the predicted and ground-truth grids, respectively, CD is defined as:(10)CD(P,G)=1|P|∑p∈Pming∈G‖p−g‖22+1|G|∑g∈Gminp∈P‖g−p‖22

Chamfer Distance evaluates bidirectional nearest-neighbor consistency and captures geometric deviations that influence surface accuracy even when they do not significantly affect the overall voxel overlap. By jointly reporting IoU, F-score, and Chamfer Distance, a comprehensive assessment of volumetric reconstruction accuracy, structural precision, and geometric consistency is provided.

To evaluate its performance, SWAGE-3D is compared with several representative 3D reconstruction methods, including 3D-R2N2 [30], which employs Recurrent Neural Networks (RNNs) for single- or multi-view reconstruction; 3D-VAE-GAN [28], a hybrid VAE–GAN framework for volumetric generation; Differentiable Ray Consistency (DRC) [65], which leverages multi-view supervision; Pix2Mesh [76], a mesh-based generative model; and Occupancy Networks (ONet) [49], which represent 3D shapes as continuous implicit functions.

In addition, we include more recent voxel-based approaches such as TMVNet [40], which utilizes a Transformer-based 3D encoder for modeling long-range volumetric dependencies, and Pix2Vox-A [39], which adopts a context-aware multi-view voxel refinement strategy.

Table 3 provides a detailed comparison of SWAGE-3D with these methods based on category-specific and average Intersection over Union (IoU) scores evaluated on the ShapeNet dataset. Although TMVNet and Pix2Vox-A achieve higher overall IoU values on ShapeNet, it is important to contextualize these results within architectural and training paradigm differences. Both TMVNet and Pix2Vox-A are designed around multi-view supervision and feature fusion mechanisms, enabling stronger geometric consistency during training. In contrast, SWAGE-3D operates under a single-view generative framework enhanced with depth integration and adversarial learning.

TMVNet leverages a Transformer-based 3D encoder to model long-range volumetric dependencies across multiple views, which provides stronger structural priors during reconstruction. Similarly, Pix2Vox-A employs a context-aware refinement strategy that explicitly aggregates multi-view voxel predictions, leading to improved geometric completeness.

By comparison, SWAGE-3D prioritizes generative modeling stability and distribution learning through a hybrid VAE-WGAN framework. While adversarial learning enhances structural realism and diversity, it does not directly optimize IoU in a purely supervised regression sense. Therefore, the slightly lower IoU values should be interpreted as a trade-off between generative robustness and direct voxel regression performance rather than a deficiency in reconstruction capability.

The results reveal a competitive landscape where different methods excel in specific categories. Across categories, SWAGE-3D consistently improves over the baseline 3D-VAE-GAN and achieves its most pronounced gains on rigid object classes (e.g., Car and Telephone), suggesting that the proposed stabilization and depth fusion improve reconstruction reliability. For instance, among single-view baselines, SWAGE-3D demonstrates a substantial margin in the Car category, achieving an IoU score of 0.835, and remains highly competitive even against multi-view methods. This exceptional performance in rigid, man-made categories is largely driven by the depth-aware encoder, which successfully resolves the depth ambiguity of flat surfaces that plagues purely RGB-based moels. While multi-view architectures like TMVNet establish the upper bound in categories such as Bench, Cabinet, and Rifle, SWAGE-3D achieves the highest average IoU among the compared single-view generative voxel-based baselines. This underscores the framework’s ability to deliver high-quality reconstructions across a wide range of object geometries, offering a robust balance between detail and structural coherence.

It is worth noting that while implicit representation methods like ONet excel in organic shapes due to their continuous nature, SWAGE-3D demonstrates superior performance in rigid, geometric objects (e.g., Cars with 0.835 IoU vs. ONet’s 0.737). This indicates that our voxel-based approach, fortified with depth guidance and attention, is particularly effective for preserving the structural integrity of man-made objects with defined planar surfaces.

While the proposed SWAGE-3D model demonstrates superior performance in volumetric objects such as cars and loudspeakers, a slight performance drop is observed in categories characterized by thin and complex structures, such as Rifle and Bench. This limitation is attributed to the fixed voxel resolution (323) and the regularization effects of spectral normalization, which may occasionally treat fine structural details as high-frequency noise during the reconstruction process. Increasing the voxel resolution in future works could effectively mitigate this limitation.

To further validate the effectiveness of SWAGE-3D, we compare its performance with the baseline 3DVAEGAN model using the IoU metric, as illustrated in Fig. 6. When evaluating individual objects, SWAGE-3D demonstrates significant improvements. Specifically, it achieves a 15.2% higher IoU for the Airplane category and a 19.4% improvement for the Table category compared to 3DVAEGAN. For objects like Display and Telephone, SWAGE-3D shows modest gains of 1.9% and 5.1%, respectively. However, in the Rifle category, 3DVAEGAN outperforms SWAGE-3D by 7.7%. Compared to the baseline 3D-VAE-GAN evaluated under the same single-view voxel reconstruction protocol, SWAGE-3D improves the average IoU from 0.491 to 0.573, corresponding to a 16.7% relative gain.

Fig. 7 provides qualitative insights into the performance of SWAGE-3D. The figure illustrates the input RGB image, the associated depth map (colorized for enhanced visibility), the voxelized model generated by SWAGE-3D, the corresponding ground truth voxelized model from the dataset, and the intersection volume between the predicted and ground truth voxels. To highlight performance variations, examples representing the highest (Car), average (Airplane), and lowest (Bench) IoU values achieved by SWAGE-3D are included. These qualitative results corroborate the quantitative findings, demonstrating the framework’s ability to produce highly accurate and structurally coherent 3D models.

Fig. 8 provides a qualitative and quantitative comparison of 3D-VAE-GAN, 3D-R2N2, Pix2Mesh, and SWAGE-3D. Depth maps are included in the inputs section but are only utilized for SWAGE-3D. Ground Truth refers to the actual 3D voxelized models in the dataset. Analysis of the 3D results highlights the superior performance of SWAGE-3D. Similar to the qualitative results presented earlier, SWAGE-3D demonstrates strong performance for Airplane, Car, Lamp, Loudspeaker, Table, and Telephone. For the Airplane, it produces a model very similar to the real one, though minor extraneous voxels appear outside the structure. For the Bench, while there are deficiencies in the railing parts, the basic skeleton is correctly constructed.

For the Cabinet, despite some issues with the arm sections, the fundamental structure is accurately produced. For the Car, the generated model closely resembles the real object. For the Chair, errors exist in the backrest and seating areas, but the overall production is successful. For the Display, the model includes the stand and is well-produced. The Lamp is reconstructed with high success. For the Loudspeaker, despite minor errors in the front section, the output is satisfactory. For the Rifle, intermediate voxels are missing. This limitation is attributed to the fixed voxel resolution (323) and the aggressive regularization of Spectral Normalization, which occasionally treats very thin structures as high-frequency noise. However, the model still correctly localizes the main body, suggesting that higher resolutions in future work could resolve this. The Sofa is produced with great success, with only minor voxel deficiencies in the seating area. For the Table, the generated model closely matches the real one, though minor errors persist in the foot sections. The Telephone is reconstructed with 100% accuracy. For the Watercraft, the model performs well, despite some deficiencies in the front and upper sections. Overall, SWAGE-3D achieves excellent results compared to other models in the literature, particularly excelling in specific object categories.

To systematically quantify the individual contributions of the integrated techniques within the SWAGE-3D framework, an incremental ablation study was conducted. Due to the high computational demand of full-scale training, this study was performed on six representative categories (Airplane, Bench, Car, Lamp, Loudspeaker, and Telephone) that exhibit diverse geometric characteristics. To ensure a fair and rigorous evaluation, all ablation variants were trained under a separate, strictly unified experimental setup to isolate the relative gain of each component. While absolute values may vary slightly from the peak performance results reported in Table 3 due to the stochastic nature of adversarial training, the performance trends across configurations remain entirely consistent.

Five model variants were evaluated: (1) a baseline 3D-VAE-GAN using 3-channel RGB input, (2) the integration of monocular depth maps (RGB-D), (3) the addition of Wasserstein GAN loss with Gradient Penalty (WGAN-GP), (4) the inclusion of Self-Attention modules, and (5) the final weighted ensemble inference. The results across Mean IoU, F-Score, and Chamfer Distance (CD) are summarized in Tables 4–6, respectively.

The systematic ablation reveals that the integration of depth maps (Config 2) is the primary driver for recovering geometric fidelity, resolving ambiguities where RGB data alone is insufficient. This is most evident in the Lamp category, where depth maps facilitated a substantial reduction in total geometric error. Subsequent architectural refinements via WGAN-GP and Self-Attention (Configs 3 and 4) focused on enhancing global shape consistency and boundary sharpness. While point-wise metrics like Mean IoU exhibit minor fluctuations, a common characteristic of GAN-based models where distributional realism is prioritized over local MSE, the structural metrics (F-Score and CD) indicate a consistent downward trend in geometric error across configurations. This study confirms that the cumulative success of SWAGE-3D arises from the synergistic combination of depth-assisted geometric initialization, stabilized adversarial learning, and robust ensemble inference.

To address the scalability of SWAGE-3D and evaluate its performance at higher voxel densities, we conducted a comparative analysis between the standard 323 resolution and a higher 643 grid size. This experiment was specifically performed on the Bench category, as it contains thin and complex structures that provide a rigorous test for high-resolution 3D generation. To ensure a fair assessment of computational efficiency and scaling behavior, both models were trained for 50 epochs under consistent experimental conditions.

The results, summarized in Table 7, demonstrate that SWAGE-3D effectively scales to higher resolutions while maintaining structural integrity. While the 643 resolution significantly increases the learning complexity and the state space of the voxel grid, the model successfully captures the essential geometric properties of the target shapes. As expected, the computational demand increases with resolution; specifically, GPU memory usage rose from 0.41 to 1.34 GB, and the training time per epoch increased from 1.08 to 11.25 min. Although the IoU and F-score values for the 643 resolution are slightly lower than those of 323 within the fixed 50-epoch training budget the model’s ability to handle eight times the voxel density confirms its architectural robustness and scalability for high-fidelity 3D reconstruction.

The experimental findings provide a comprehensive assessment of the proposed SWAGE-3D framework across reconstruction accuracy, geometric consistency, stabilization behavior, and scalability. The results collectively indicate that the integration of depth-assisted encoding and stabilized adversarial training contributes to improved reconstruction robustness under a single-view voxel-based setting.

The quantitative comparison on ShapeNet demonstrates that SWAGE-3D consistently outperforms the baseline 3D-VAE-GAN across multiple object categories in terms of IoU. The most pronounced improvements are observed in rigid and structurally well-defined categories such as Car and Telephone, suggesting that depth-guided encoding effectively mitigates depth ambiguity in planar and symmetric structures. This indicates that the RGB-D fusion strategy plays a central role in improving geometric reliability when only a single image is available.

The incremental ablation study further clarifies the relative contributions of architectural components. The depth map integration yields the most substantial improvement in mean IoU and F-score, confirming that geometric cues extracted from monocular depth estimation significantly enhance volumetric inference. The inclusion of WGAN-GP and spectral normalization primarily contributes to training stabilization and distribution alignment rather than dramatic IoU gains. While the improvement in volumetric overlap metrics is moderate, the reduction in Chamfer Distance suggests improved geometric smoothness and surface coherence. Similarly, the self-attention module provides marginal gains in voxel overlap but contributes to enhanced structural consistency, particularly reflected in distance-based evaluation. The ensemble strategy offers limited incremental improvement in IoU but helps reduce variance and improve robustness during inference.

Taken together, these findings suggest that performance gains do not stem from a single dominant architectural modification alone, but rather from a controlled combination of complementary stabilization and geometric enhancement strategies. The improvements are more pronounced in structural reliability and geometric consistency than in raw volumetric overlap alone.

The inclusion of F-score and Chamfer Distance provides a more nuanced understanding of reconstruction behavior. While IoU captures volumetric overlap, F-score reflects precision-recall balance of occupancy prediction, and Chamfer Distance quantifies geometric proximity. The observed trend, where some configurations yield modest IoU changes but measurable Chamfer improvements, indicates that adversarial stabilization and attention mechanisms primarily refine surface quality rather than drastically altering voxel occupancy ratios. This highlights the importance of multi-metric evaluation when analyzing generative volumetric models.

The resolution analysis comparing 32323 and 64643 voxel grids reveals a critical trade-off between geometric granularity and computational feasibility. Although the 64643 configuration increases representational capacity, it leads to reduced IoU and F-score under the current hardware constraints, while significantly increasing memory consumption and training time. The cubic growth in voxel space dramatically enlarges the optimization landscape, making stable adversarial training more challenging on a single GTX 980 Ti GPU. These findings indicate that while the architecture is technically scalable to higher resolutions, effective optimization at 64643 resolution requires either stronger regularization strategies or more advanced computational resources. Therefore, 32323 remains a practical and balanced resolution choice for stable single-view adversarial voxel reconstruction under moderate hardware settings.

Despite its improvements, SWAGE-3D has several limitations. First, the voxel-based representation inherently suffers from discretization artifacts and cubic computational complexity. Although adversarial learning improves structural realism, thin structures and fine details remain difficult to reconstruct at low resolutions. While higher resolutions are theoretically feasible, they introduce substantial computational overhead and optimization instability. Second, the model relies on monocular depth estimation as a preprocessing step. Errors in predicted depth maps directly propagate to the latent representation, potentially limiting reconstruction fidelity in visually ambiguous or textureless regions.

Third, while the incremental ablation study demonstrates the contribution of each component, the magnitude of improvement beyond depth integration is moderate. This suggests that geometric cues play a more dominant role than adversarial refinement in single-view voxel reconstruction under the current configuration. Finally, SWAGE-3D is evaluated under a single-view generative paradigm. Multi-view transformer-based methods achieve higher absolute IoU values by leveraging stronger geometric supervision. The proposed framework prioritizes stabilization and distribution-aware volumetric generation rather than purely supervised regression optimization.

Future research may explore hybrid voxel-implicit representations to alleviate discretization constraints, adaptive resolution strategies to dynamically allocate voxel density, and more advanced regularization techniques to stabilize higher-resolution adversarial training. Additionally, integrating uncertainty-aware depth estimation or jointly optimizing depth and voxel generation in an end-to-end manner may further enhance geometric consistency.