EdgeFusion: A Diffusion Framework for Real-Time 3D Generation on Resource-Constrained Devices

EdgeFusion: A Diffusion Framework for Real-Time 3D Generation on Resource-Constrained Devices | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article EdgeFusion: A Diffusion Framework for Real-Time 3D Generation on Resource-Constrained Devices Nnaemeka Kingsley Ugwumba This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8168090/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This research tackles a big problem in 3D artificial intelligence: the models that create 3D objects are usually too large and slow to run on small devices like phones or embedded systems. We present EdgeFusion, a new method that makes these 3D AI models much smaller and faster. Our approach combines several smart shrinking techniques, like removing unnecessary parts of the model and simplifying its calculations. The results are very strong. We reduced the model size by 97%, from 5.58 megabytes to just 0.17 megabytes. We also made it 3.9 times faster. This proves that it is possible to run powerful 3D generation on devices with limited resources, opening new doors for mobile and edge computing applications. Artificial Intelligence and Machine Learning 3D Generation Edge Computing AI Optimization Model Compression Efficient AI Diffusion Models Tiny Machine Learning Mobile AI 3D Artificial Intelligence Lightweight Models Figures Figure 1 Figure 2 1. Introduction The creation of 3D objects and scenes by artificial intelligence has seen amazing progress. Systems can now generate realistic 3D models from simple text descriptions. However, these powerful AI models come with a major problem: they are incredibly demanding on computer resources. They require strong processors and a lot of memory, which means they can only run on powerful cloud servers or expensive desktop computers. This makes them unusable for most real world applications on everyday devices. Imagine a doctor using a tablet to generate a 3D model of a patient's organ for planning surgery, or an architect creating 3D building designs on a mobile device at a construction site. These scenarios are impossible with today's large 3D AI models. The challenge is that edge devices like phones, tablets, and embedded systems have very limited processing power, small memory, and strict battery life constraints. This creates a significant gap between what 3D AI can do and where it can be used. This research introduces EdgeFusion, a new framework designed to close this gap. Our goal is to make 3D generation practical for resource constrained devices without losing the core ability to create useful 3D content. We achieve this by fundamentally rethinking how these AI models are built and optimized. The main objectives of this work are: To build a standard 3D diffusion model as a baseline and understand its performance characteristics. This means creating a working AI system that can generate simple 3D shapes and carefully measuring its size, speed, and memory usage to establish a performance starting point. To design and test multiple model optimization techniques specifically for 3D tasks. We will explore three main methods: making the model's architecture simpler by removing layers, setting tiny unimportant weights to zero, and reducing the numerical precision of calculations. To develop a combined optimization strategy called EdgeFusion that integrates all effective techniques. We will create a single, highly efficient model that uses architectural changes, weight reduction, and numerical simplification together to achieve maximum performance gains. To thoroughly evaluate the optimized models and demonstrate their practical value. We will compare our EdgeFusion model against the original baseline across multiple metrics including model size, inference speed, number of parameters, and generation quality to prove its effectiveness for edge deployment. By achieving these objectives, this work provides a clear path forward for deploying advanced 3D AI capabilities on the devices where they are most needed, ultimately making 3D generation more accessible and practical for real world applications. 2. Related Work This research builds upon established work in three-dimensional generative modeling, model compression techniques, and efficient computing for edge devices. This section surveys the relevant literature and identifies the research gap that this work addresses. 2.1 Three-Dimensional Generative Models The field of 3D content generation has evolved significantly from early voxel-based approaches (Wu et al., 2016) to more sophisticated implicit representation methods. The recent success of diffusion models in image generation (Ho et al., 2020) has inspired their adaptation to 3D domains. Zhou et al. (2021) demonstrated the application of diffusion processes for 3D point cloud generation, while Luo et al. (2022) extended these principles to neural radiance fields for novel view synthesis. These approaches have shown remarkable capability in generating high-quality 3D content, but they share a common limitation: substantial computational requirements that necessitate powerful GPU infrastructure and extensive memory resources, making them impractical for resource-constrained environments. 2.2 Model Compression and Optimization Model compression techniques have emerged as essential tools for deploying deep learning models in practical applications. Pruning methods, both unstructured (Han et al., 2015) and structured (Li et al., 2017), have demonstrated effective model size reduction while maintaining functionality. Quantization approaches, particularly integer quantization (Jacob et al., 2018), have enabled significant memory footprint reduction and acceleration on supported hardware. Knowledge distillation (Hinton et al., 2015) has provided a framework for transferring knowledge from large teacher models to compact student networks. While these techniques have proven successful in various 2D computer vision tasks (He et al., 2018) and natural language processing applications (Devlin et al., 2019), their systematic application and evaluation specifically for 3D generative tasks remains limited in the current literature. 2.3 Efficient Computing for Edge Devices The growing demand for on-device artificial intelligence has catalyzed the tinyML paradigm (David et al., 2021), focusing on machine learning deployment under severe resource constraints. Research in this area has primarily concentrated on classification and regression tasks for applications such as mobile health monitoring (Ravi et al., 2022) and industrial IoT systems (Zhang et al., 2020). The computational characteristics of generative tasks, particularly those involving three-dimensional content creation, present fundamentally different challenges compared to discriminative tasks. The iterative sampling process inherent in diffusion models, combined with the high-dimensional nature of 3D representations, creates unique computational demands that conventional edge AI optimization strategies are not designed to address. 2.4 Research Gap and Contribution While significant progress has been made in individual domains of 3D generation, model compression, and edge computing, the integration of these research streams remains largely unexplored. Current literature lacks comprehensive studies on optimizing 3D diffusion models specifically for resource-constrained environments. No existing work provides a systematic framework that combines architectural efficiency, parameter reduction, and numerical optimization tailored for 3D generative tasks on edge devices. This work addresses this gap by proposing the EdgeFusion framework, which represents a novel approach to making 3D diffusion models practical for deployment on devices with limited computational resources, thereby enabling new applications in mobile and embedded systems. 3. Methodology This section details the comprehensive methodology developed for the EdgeFusion framework. The approach encompasses dataset preparation, baseline model development, and the implementation of multiple optimization techniques specifically designed for 3D diffusion models on edge devices. 3.1 Dataset and Preprocessing The ModelNet40 dataset was utilized for this research, comprising 12,311 3D CAD models across 40 object categories. The dataset was processed through a standardized pipeline to ensure consistency and compatibility with edge device constraints. Each 3D model was converted to a 32×32×32 voxel grid representation using a custom voxelization algorithm. This resolution was selected as it provides sufficient geometric detail while maintaining computational feasibility for edge deployment. The voxelization process involved normalizing all models to a unit cube and applying binary occupancy encoding, where each voxel element was assigned a value of 1 for occupied space and 0 for empty space. The dataset was partitioned using the standard training and testing splits, with 9,843 models allocated for training and 2,468 models for evaluation. 3.2 Baseline 3D Diffusion Model A standard 3D diffusion model was implemented as the baseline for performance comparison. The model architecture consisted of a 3D U-Net denoising network with four encoding and decoding blocks. The network employed 3D convolutional layers with kernel size 3×3×3 and progressively increasing channel dimensions of 16, 32, 64, and 128. Skip connections were integrated between corresponding encoder and decoder blocks to preserve spatial information. The diffusion process was configured with 1000 timesteps using a linear noise schedule. The model was trained using mean squared error loss between predicted and actual noise components, with the AdamW optimizer at a learning rate of 1×10⁻⁴. This baseline configuration achieved a parameter count of 1.46 million and a model size of 5.58 megabytes. 3.3 EdgeFusion Optimization Framework The EdgeFusion framework incorporates three complementary optimization strategies specifically designed for 3D diffusion models targeting edge deployment. 3.3.1 Architectural Pruning A systematic reduction of the baseline model architecture was performed to create a computationally efficient variant. The network depth was reduced from four to two encoding-decoding blocks, and the channel dimensions were scaled down to 8, 16, and 32 in the bottleneck layer. The convolutional blocks were simplified by removing redundant layers and implementing depthwise separable convolutions where appropriate. This architectural optimization resulted in a substantial reduction of parameters while maintaining the essential spatial processing capabilities required for 3D data. 3.3.2 Weight Pruning A magnitude-based pruning approach was applied to eliminate redundant parameters from the trained models. The pruning threshold was empirically determined through iterative experimentation, with weights below the 0.02 magnitude threshold set to zero. This process was conducted in a structured manner to preserve the functional integrity of the feature extraction layers. The pruning mask was applied globally across all convolutional layers, with periodic retraining cycles to recover any potential accuracy loss from the pruning operation. 3.3.3 Quantization Post-training quantization was implemented to reduce the numerical precision of model parameters and activations. The 32-bit floating-point representations were converted to 8-bit integer format using a symmetric quantization scheme. Calibration was performed using a representative subset of the training data to determine appropriate scaling factors for each layer. The quantization process focused on the convolutional layers and linear transformations within the network, with special consideration for the skip connection operations to maintain gradient flow during inference. 3.4 Implementation Details All experiments were conducted using PyTorch 2.0 with CUDA acceleration where available. The training process utilized a batch size of 4 due to memory constraints of the voxel representations. Data augmentation techniques including random rotation and scaling were applied during training to improve model robustness. The optimization techniques were implemented sequentially, with each optimization step followed by a brief fine-tuning phase to stabilize performance. The entire framework was designed with modularity in mind, allowing for independent application of each optimization technique or their combined implementation as required by specific deployment scenarios. 3.5 Evaluation Metrics The performance of all models was evaluated using multiple metrics to provide comprehensive assessment. Model size was measured in megabytes, including all parameters and necessary metadata for deployment. Inference time was measured in milliseconds using standardized input tensors on consistent hardware configurations. Parameter efficiency was calculated as the ratio of active parameters to total model capacity. Additional metrics included memory footprint during inference and computational complexity in terms of floating-point operations required for a single forward pass. These metrics collectively provided a holistic view of model efficiency and suitability for edge deployment. 4. Results and Discussion This section presents a comprehensive evaluation of the EdgeFusion framework, comparing the performance of optimized models against the baseline across multiple metrics. The results demonstrate the effectiveness of our optimization strategies for enabling 3D diffusion models on resource-constrained devices. 4.1 Experimental Setup and Evaluation Metrics All experiments were conducted on a standardized testing environment using the ModelNet40 test set comprising 2,468 3D models. Performance was evaluated using four key metrics: model size (MB), parameter count, inference time (ms), and memory footprint during operation. Inference time was measured as the average processing time for 100 consecutive forward passes using batch size 1 to simulate real-time edge deployment scenarios. 4.2 Baseline Model Performance The baseline 3D diffusion model established the performance reference point for subsequent comparisons. The model achieved functional 3D generation capability with a parameter count of 1,460,993 and a storage size of 5.58 MB. The average inference time measured 2.47 milliseconds per forward pass on the test hardware. While this baseline provided satisfactory generation quality, its computational demands rendered it unsuitable for deployment on typical edge devices with limited memory and processing capabilities. Figure 1 provides a comprehensive comparison of all model variants across four key performance metrics: model size, inference time, parameter count, and speedup factor. The visualization clearly demonstrates the progressive improvements achieved through each optimization technique. 4.3 Individual Optimization Results The evaluation of individual optimization techniques revealed distinct performance characteristics and efficiency gains. Architectural pruning emerged as the most impactful strategy, achieving a 97.0% reduction in both parameters (from 1.46M to 43.7K) and model size (from 5.58 MB to 0.17 MB) while improving inference speed by 2.7×, demonstrating that fundamental network redesign delivers the most substantial benefits for edge deployment. Weight pruning successfully sparsified the model by zeroing out 90.2% of parameters (1,315,561 weights) and achieved a 1.2× speedup, though it provided limited storage reduction due to implementation constraints that maintain the original parameter structure. Quantization delivered more modest improvements with a 1.1× speedup and theoretical 50% storage reduction potential, showing particular promise for specialized hardware environments optimized for integer operations, though its benefits in general computing scenarios were less pronounced. Table 1 presents detailed numerical metrics for all model configurations, quantifying the exact improvements in model size, parameter count, inference time, and efficiency gains relative to the baseline. Table 1 : Detailed metrics for quantitative analysis 4.3.1 Architectural Pruning Effectiveness The architectural pruning strategy yielded the most significant improvements in model efficiency. By reducing network depth and channel dimensions, the optimized architecture achieved a parameter count of 43,665, representing a 97.0% reduction compared to the baseline. The model size decreased to 0.17 MB while maintaining the core 3D generation functionality. Inference time improved to 0.91 milliseconds, representing a 2.7× speedup over the baseline. This demonstrates that careful architectural design can achieve substantial efficiency gains without compromising fundamental model capabilities. 4.3.2 Weight Pruning Performance The magnitude-based weight pruning approach successfully identified and eliminated redundant parameters while preserving model functionality. The pruning process zeroed out 1,315,561 parameters, achieving 90.2% sparsity in the weight matrices. Interestingly, while the parameter count reduction was substantial, the model size remained at 5.58 MB due to the storage format requirements. The inference time improved to 2.13 milliseconds, representing a 1.2× speedup. This result highlights that weight pruning primarily benefits computational efficiency rather than storage requirements in standard deployment scenarios. 4.3.3 Quantization Impact The 8-bit quantization implementation demonstrated moderate improvements in model efficiency. The quantized model maintained the same parameter count as the baseline but achieved a theoretical storage reduction potential of 50% (to 2.79 MB) when using integer-only arithmetic. The inference time improved to 2.27 milliseconds, representing a 1.1× speedup. The modest performance gain suggests that quantization provides greater benefits in specialized hardware environments optimized for integer operations, which are common in edge computing platforms. 4.4 EdgeFusion Combined Optimization The integrated EdgeFusion framework, combining all three optimization strategies, achieved the most compelling results for edge deployment. The final model attained a parameter count of 43,505 with a model size of 0.17 MB, representing a 97.0% reduction in both metrics compared to the baseline. The inference time reached 0.63 milliseconds, delivering a 3.9× speedup. This comprehensive optimization resulted in a model that occupies only 170 KB of storage space while maintaining the essential 3D generation capabilities required for practical applications. Figure 2 analyzes the optimization trade-offs, showing the relationship between model size and inference speed across different approaches, along with the individual impact of each optimization technique on model size reduction. 4.5 Performance Trade-off Analysis The evaluation reveals important trade-offs between the optimization techniques. Architectural pruning provided the most substantial benefits across all metrics, fundamentally redesigning the network for efficiency. Weight pruning offered computational improvements but limited storage benefits due to implementation constraints. Quantization showed potential for specialized hardware but provided modest gains in general computing environments. The combined EdgeFusion approach demonstrated that synergistic integration of these techniques can achieve near-optimal performance across all evaluation metrics. Table 2 summarizes the characteristics and effectiveness of each optimization technique, providing a clear overview of how architectural pruning, weight pruning, quantization, and their combination contribute to the final EdgeFusion performance. Table 2 : Optimization techniques summary and effectiveness 4.6 Generation Quality Assessment Qualitative analysis of generated samples confirmed that the optimized models maintained functional 3D generation capabilities despite the substantial parameter reduction. While some detail loss was observable in complex geometries, the core shape representation remained intact for most object categories. This balance between efficiency and capability makes the EdgeFusion framework particularly suitable for applications where computational constraints outweigh the need for high-frequency detail, such as mobile AR/VR applications and real-time design tools. 4.7 Discussion and Implications The results demonstrate that 3D diffusion models can be effectively optimized for edge deployment without sacrificing core functionality. The 97% model size reduction and 3.9× inference speedup achieved by EdgeFusion represent significant advancements toward practical 3D AI on resource-constrained devices. These improvements enable new application scenarios in mobile computing, embedded systems, and real-time 3D content generation that were previously limited by computational requirements. The success of computer vision systems in agricultural applications, such as Ugwumba's (2025) biomass estimation framework, reinforces our findings that well-optimized models can deliver robust performance in practical edge computing scenarios. This parallel development across domains suggests a broader trend toward making advanced AI capabilities available outside traditional high-performance computing environments Hence, the success of architectural pruning suggests that model efficiency should be considered from the initial design phase rather than as a post-training optimization. The complementary nature of the optimization techniques indicates that holistic approaches combining multiple strategies yield superior results compared to individual optimizations. This insight provides valuable guidance for future research in efficient 3D model design and deployment. The EdgeFusion framework establishes a foundation for bringing advanced 3D generation capabilities to edge devices, potentially enabling new applications in mobile gaming, augmented reality, and on-device content creation tools. The demonstrated efficiency gains make real-time 3D generation feasible on hardware previously considered insufficient for such tasks, expanding the accessibility and applicability of 3D AI technologies. 5. Conclusion and Future Work This research has successfully demonstrated that significant optimization of 3D diffusion models for edge deployment is not only possible but can be achieved with remarkable efficiency. The proposed EdgeFusion framework has addressed the critical challenge of computational intensity in 3D generative AI by systematically implementing and combining architectural pruning, weight sparsification, and quantization techniques. The results conclusively show that our approach reduces model size by 97.0%, decreases parameter count from 1.46 million to 43.5 thousand, and achieves a 3.9× inference speedup while maintaining essential 3D generation capabilities. These achievements represent a substantial advancement toward making sophisticated 3D content generation accessible on resource-constrained devices, thereby bridging the gap between state-of-the-art generative AI and practical edge computing applications. Looking forward, the success of reinforcement learning approaches in optimization tasks, such as the Deep Q-Network for task prioritization by Ugwumba & Jaja ( 2025 ), suggests promising directions for enhancing our EdgeFusion framework. Future iterations could employ similar reinforcement learning techniques to automatically discover optimal model compression strategies that dynamically balance the trade-offs between model size, inference speed, and generation quality. The implications of this work extend across multiple domains, including mobile augmented reality, real-time design tools, and embedded 3D visualization systems. By reducing the computational barriers to 3D generative AI, EdgeFusion enables new applications where processing power, memory availability, and energy consumption are limiting factors. The framework's modular design allows for flexible adaptation to various deployment scenarios, from smartphones to specialized edge hardware. For future work, several promising directions emerge. First, investigating the application of EdgeFusion to more complex 3D representations beyond voxel grids, such as neural radiance fields or implicit neural representations, could further expand its utility. Second, exploring hardware-aware neural architecture search could automate and optimize the model design process for specific edge platforms. Third, developing dynamic optimization strategies that adapt model complexity based on available resources would enhance practical deployment in variable computing environments. Finally, extending the framework to support conditional generation and multi-modal inputs would increase its applicability to real-world scenarios requiring interactive and context-aware 3D content creation. Declarations Ethical Approval Not applicable. This research did not involve human participants, animal subjects, or any primary data collection from living entities. Competing Interests The authors declare no competing interests, financial or non-financial, relevant to the content of this article. Funding The authors received no specific funding for this work. Authorship Contribution Nnaemeka KIngsley Ugwumba: Conceptualization, Methodology, Software, Writing - Original Draft. The author reviewed and approved the final manuscript. Data Availability Declaration All data generated or analysed during this study, including the figures and source code, are available in the following GitHub repository https://github.com/KingsleyTechie/EdgeFusion3D References Balraj. (2019). ModelNet40 [Data set]. Kaggle. https://www.kaggle.com/datasets/balraj98/modelnet40-princeton-3d-object-dataset David, R., Duke, J., Jain, A., Reddi, V. J., Jeffries, N., Li, J., Kreeger, N., Nappier, I., Natraj, M., Regev, S., Rhodes, R., Wang, T., & Warden, P. (2021). TensorFlow Lite Micro: Embedded machine learning for TinyML. Journal of Machine Learning Research, 22(1), 1-24. https://doi.org/10.5555/3454287.3454456 Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 1, 4171–4186. https://doi.org/10.18653/v1/N19-1423 Han, S., Pool, J., Tran, J., & Dally, W. J. (2015). Learning both weights and connections for efficient neural networks. Advances in Neural Information Processing Systems, 28, 1135–1143. https://doi.org/10.5555/2969239.2969364 He, Y., Lin, J., Liu, Z., Wang, H., Li, L.-J., & Han, S. (2018). AMC: AutoML for model compression and acceleration on mobile devices. Proceedings of the European Conference on Computer Vision (ECCV), 784–800. https://doi.org/10.1007/978-3-030-01228-1_47 Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531. https://doi.org/10.48550/arXiv.1503.02531 Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840–6851. https://doi.org/10.5555/3495724.3496298 Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., & Kalenichenko, D. (2018). Quantization and training of neural networks for efficient integer-arithmetic-only inference. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2704–2713. https://doi.org/10.1109/CVPR.2018.00286 Li, H., Kadav, A., Durdanovic, I., Samet, H., & Graf, H. P. (2017). Pruning filters for efficient ConvNets. International Conference on Learning Representations. https://doi.org/10.48550/arXiv.1608.08710 Luo, X., & Liu, Z. (2022). Diffusion-based 3D scene synthesis with neural radiance fields. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4168–4178. https://doi.org/10.1109/CVPR52688.2022.00414 Ravi, D., Wong, C., Lo, B., & Yang, G.-Z. (2022). Deep learning for human activity recognition: A resource-efficient implementation on edge devices. IEEE Transactions on Mobile Computing, 21(4), 1234–1248. https://doi.org/10.1109/TMC.2020.3034211 Ugwumba, N. K. (2025). Computer vision for pasture biomass estimation: Enabling data-driven grazing decisions through multi-modal deep learning. Research Square. https://doi.org/10.21203/rs.3.rs-8071124/v1 Ugwumba, N. K., & Jaja, P. S. (2025). Enhanced task prioritization system using Deep-Q-Network model. International Journal of Computer Science Engineering Techniques, 9(6), IJCSE-V9I6P15. https://doi.org/10.5281/zenodo.17636107 Wu, J., Zhang, C., Xue, T., Freeman, B., & Tenenbaum, J. (2016). Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. Advances in Neural Information Processing Systems, 29, 82–90. https://doi.org/10.5555/3157096.3157106 Zhang, C., Patras, P., & Haddadi, H. (2020). Deep learning in mobile and wireless networking: A survey. IEEE Communications Surveys & Tutorials, 22(3), 2224–2287. https://doi.org/10.1109/COMST.2019.2904897 Zhou, L., Du, Y., & Wu, J. (2021). 3D shape generation and completion through point-voxel diffusion. Proceedings of the IEEE/CVF International Conference on Computer Vision, 5826–5835. https://doi.org/10.1109/ICCV48922.2021.00578 Tables Table 1 and 2 are available in the Supplementary Files section. Additional Declarations The authors declare no competing interests. Supplementary Files Table1.png Table2.png Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8168090","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":548380683,"identity":"b8a48562-060f-4cfc-9a32-540884ff4a6d","order_by":0,"name":"Nnaemeka Kingsley Ugwumba","email":"data:image/png;base64,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","orcid":"https://orcid.org/0009-0000-2493-9846","institution":"University of Port Harcourt, Port Harcourt, Nigeria","correspondingAuthor":true,"prefix":"","firstName":"Nnaemeka","middleName":"Kingsley","lastName":"Ugwumba","suffix":""}],"badges":[],"createdAt":"2025-11-20 22:48:37","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":true,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":true},"doi":"10.21203/rs.3.rs-8168090/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8168090/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96711062,"identity":"b43cfe09-c6d5-4911-bdd9-79718cc26199","added_by":"auto","created_at":"2025-11-25 10:11:34","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":405014,"visible":true,"origin":"","legend":"","description":"","filename":"EdgeFusion.docx","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/c692a79e4b76608bbb9009a6.docx"},{"id":96697216,"identity":"76db4e5e-6634-4e6e-ac3d-191aa58fa06d","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8168090.json","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/ebf1d6d93b13c03bdffe558b.json"},{"id":96697217,"identity":"c7fb7b33-eeb3-4e78-ad25-fe29cc5290e8","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56295,"visible":true,"origin":"","legend":"","description":"","filename":"rs81680900enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/9915bccd06cd2933697bd307.xml"},{"id":96697221,"identity":"abd10efb-c86b-4a46-8cb2-937083d860f3","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":139069,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/1f3a984aa6daafb1458642a8.png"},{"id":96697218,"identity":"7268a4e0-2d57-4182-bc5a-953be66cbd0e","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":80798,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/4bca62cb9fe4c29361445b61.png"},{"id":96711367,"identity":"73127b2c-6174-4172-be47-6031bf14e4ca","added_by":"auto","created_at":"2025-11-25 10:11:54","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":82040,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/47dd78f45ad3f1b5269440a4.png"},{"id":96697222,"identity":"69fc79c3-b65b-461c-ae50-40e6bd230683","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":72066,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/1c4dab7ea83f9344ecc4f49f.png"},{"id":96711235,"identity":"2f11314d-2004-44b7-9735-4b0f8f0e0954","added_by":"auto","created_at":"2025-11-25 10:11:47","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35998,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/78f9bf41c880a4671b263ab3.png"},{"id":96697223,"identity":"873e7ad6-0187-40ed-946a-d6e9184e220b","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":19452,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/8d632453b512783f713cff81.png"},{"id":96711268,"identity":"55549cd8-1843-4420-ae45-c313457caa86","added_by":"auto","created_at":"2025-11-25 10:11:49","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":19902,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/8829c650d81dcf5e6dc4ec62.png"},{"id":96711489,"identity":"4565d6c9-b6e3-4d29-a4de-a814514861a1","added_by":"auto","created_at":"2025-11-25 10:12:05","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17283,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/e1c2361c98499a0967545768.png"},{"id":96697226,"identity":"94223d9f-9830-4225-9aad-e95c1914cac0","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":55600,"visible":true,"origin":"","legend":"","description":"","filename":"rs81680900structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/b1a58d8922cac8dae66da6d7.xml"},{"id":96697228,"identity":"dc9b6d4d-12b0-44ef-862f-0f951d4937d4","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"html","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":60685,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/281a9a3a63dd4019c34fbfec.html"},{"id":96697214,"identity":"67e49095-0ef3-4848-9aaa-1b7517e36af6","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":139173,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance comparison across all model variants\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/8dea38aeb00caa556517ae2a.png"},{"id":96710630,"identity":"c29675b8-6b00-4e40-b931-8f2f5425a21f","added_by":"auto","created_at":"2025-11-25 10:11:00","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":166135,"visible":true,"origin":"","legend":"\u003cp\u003eOptimization trade-offs and impact assessment\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/84ad3e379adb50dcd56283f5.png"},{"id":96712916,"identity":"54445d79-d367-4b3d-8ccf-0c0e434229af","added_by":"auto","created_at":"2025-11-25 10:17:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":912933,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/c540f616-6d93-482e-9905-c186d2508238.pdf"},{"id":96697212,"identity":"275f2e5c-72de-43be-a0a5-d68532a13a49","added_by":"auto","created_at":"2025-11-25 07:44:41","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":166540,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/0f89d6e1dbaf0e689a3811f5.png"},{"id":96710798,"identity":"70ae5860-c896-41ca-b49a-3b8a4553ac94","added_by":"auto","created_at":"2025-11-25 10:11:11","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":156496,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.png","url":"https://assets-eu.researchsquare.com/files/rs-8168090/v1/7839a2b276be96a51714db9a.png"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eEdgeFusion: A Diffusion Framework for Real-Time 3D Generation on Resource-Constrained Devices\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe creation of 3D objects and scenes by artificial intelligence has seen amazing progress. Systems can now generate realistic 3D models from simple text descriptions. However, these powerful AI models come with a major problem: they are incredibly demanding on computer resources. They require strong processors and a lot of memory, which means they can only run on powerful cloud servers or expensive desktop computers. This makes them unusable for most real world applications on everyday devices.\u003c/p\u003e\n\u003cp\u003eImagine a doctor using a tablet to generate a 3D model of a patient\u0026apos;s organ for planning surgery, or an architect creating 3D building designs on a mobile device at a construction site. These scenarios are impossible with today\u0026apos;s large 3D AI models. The challenge is that edge devices like phones, tablets, and embedded systems have very limited processing power, small memory, and strict battery life constraints. This creates a significant gap between what 3D AI can do and where it can be used.\u003c/p\u003e\n\u003cp\u003eThis research introduces EdgeFusion, a new framework designed to close this gap. Our goal is to make 3D generation practical for resource constrained devices without losing the core ability to create useful 3D content. We achieve this by fundamentally rethinking how these AI models are built and optimized.\u003c/p\u003e\n\u003cp\u003eThe main objectives of this work are:\u003c/p\u003e\n\u003col style=\"list-style-type: lower-roman;\"\u003e\n \u003cli\u003eTo build a standard 3D diffusion model as a baseline and understand its performance characteristics. This means creating a working AI system that can generate simple 3D shapes and carefully measuring its size, speed, and memory usage to establish a performance starting point.\u003c/li\u003e\n \u003cli\u003eTo design and test multiple model optimization techniques specifically for 3D tasks. We will explore three main methods: making the model\u0026apos;s architecture simpler by removing layers, setting tiny unimportant weights to zero, and reducing the numerical precision of calculations.\u003c/li\u003e\n \u003cli\u003eTo develop a combined optimization strategy called EdgeFusion that integrates all effective techniques. We will create a single, highly efficient model that uses architectural changes, weight reduction, and numerical simplification together to achieve maximum performance gains.\u003c/li\u003e\n \u003cli\u003eTo thoroughly evaluate the optimized models and demonstrate their practical value. We will compare our EdgeFusion model against the original baseline across multiple metrics including model size, inference speed, number of parameters, and generation quality to prove its effectiveness for edge deployment.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eBy achieving these objectives, this work provides a clear path forward for deploying advanced 3D AI capabilities on the devices where they are most needed, ultimately making 3D generation more accessible and practical for real world applications.\u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cp\u003eThis research builds upon established work in three-dimensional generative modeling, model compression techniques, and efficient computing for edge devices. This section surveys the relevant literature and identifies the research gap that this work addresses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1 Three-Dimensional Generative Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe field of 3D content generation has evolved significantly from early voxel-based approaches (Wu et al., 2016) to more sophisticated implicit representation methods. The recent success of diffusion models in image generation (Ho et al., 2020) has inspired their adaptation to 3D domains. Zhou et al. (2021) demonstrated the application of diffusion processes for 3D point cloud generation, while Luo et al. (2022) extended these principles to neural radiance fields for novel view synthesis. These approaches have shown remarkable capability in generating high-quality 3D content, but they share a common limitation: substantial computational requirements that necessitate powerful GPU infrastructure and extensive memory resources, making them impractical for resource-constrained environments.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2 Model Compression and Optimization\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModel compression techniques have emerged as essential tools for deploying deep learning models in practical applications. Pruning methods, both unstructured (Han et al., 2015) and structured (Li et al., 2017), have demonstrated effective model size reduction while maintaining functionality. Quantization approaches, particularly integer quantization (Jacob et al., 2018), have enabled significant memory footprint reduction and acceleration on supported hardware. Knowledge distillation (Hinton et al., 2015) has provided a framework for transferring knowledge from large teacher models to compact student networks. While these techniques have proven successful in various 2D computer vision tasks (He et al., 2018) and natural language processing applications (Devlin et al., 2019), their systematic application and evaluation specifically for 3D generative tasks remains limited in the current literature.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.3 Efficient Computing for Edge Devices\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe growing demand for on-device artificial intelligence has catalyzed the tinyML paradigm (David et al., 2021), focusing on machine learning deployment under severe resource constraints. Research in this area has primarily concentrated on classification and regression tasks for applications such as mobile health monitoring (Ravi et al., 2022) and industrial IoT systems (Zhang et al., 2020). The computational characteristics of generative tasks, particularly those involving three-dimensional content creation, present fundamentally different challenges compared to discriminative tasks. The iterative sampling process inherent in diffusion models, combined with the high-dimensional nature of 3D representations, creates unique computational demands that conventional edge AI optimization strategies are not designed to address.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.4 Research Gap and Contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWhile significant progress has been made in individual domains of 3D generation, model compression, and edge computing, the integration of these research streams remains largely unexplored. Current literature lacks comprehensive studies on optimizing 3D diffusion models specifically for resource-constrained environments. No existing work provides a systematic framework that combines architectural efficiency, parameter reduction, and numerical optimization tailored for 3D generative tasks on edge devices. This work addresses this gap by proposing the EdgeFusion framework, which represents a novel approach to making 3D diffusion models practical for deployment on devices with limited computational resources, thereby enabling new applications in mobile and embedded systems.\u003c/p\u003e"},{"header":"3. Methodology","content":"\u003cp\u003eThis section details the comprehensive methodology developed for the EdgeFusion framework. The approach encompasses dataset preparation, baseline model development, and the implementation of multiple optimization techniques specifically designed for 3D diffusion models on edge devices.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Dataset and Preprocessing\u003c/h2\u003e\u003cp\u003eThe ModelNet40 dataset was utilized for this research, comprising 12,311 3D CAD models across 40 object categories. The dataset was processed through a standardized pipeline to ensure consistency and compatibility with edge device constraints. Each 3D model was converted to a 32\u0026times;32\u0026times;32 voxel grid representation using a custom voxelization algorithm. This resolution was selected as it provides sufficient geometric detail while maintaining computational feasibility for edge deployment. The voxelization process involved normalizing all models to a unit cube and applying binary occupancy encoding, where each voxel element was assigned a value of 1 for occupied space and 0 for empty space. The dataset was partitioned using the standard training and testing splits, with 9,843 models allocated for training and 2,468 models for evaluation.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Baseline 3D Diffusion Model\u003c/h2\u003e\u003cp\u003eA standard 3D diffusion model was implemented as the baseline for performance comparison. The model architecture consisted of a 3D U-Net denoising network with four encoding and decoding blocks. The network employed 3D convolutional layers with kernel size 3\u0026times;3\u0026times;3 and progressively increasing channel dimensions of 16, 32, 64, and 128. Skip connections were integrated between corresponding encoder and decoder blocks to preserve spatial information. The diffusion process was configured with 1000 timesteps using a linear noise schedule. The model was trained using mean squared error loss between predicted and actual noise components, with the AdamW optimizer at a learning rate of 1\u0026times;10⁻⁴. This baseline configuration achieved a parameter count of 1.46\u0026nbsp;million and a model size of 5.58 megabytes.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.3 EdgeFusion Optimization Framework\u003c/h2\u003e\u003cp\u003eThe EdgeFusion framework incorporates three complementary optimization strategies specifically designed for 3D diffusion models targeting edge deployment.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\u003ch2\u003e3.3.1 Architectural Pruning\u003c/h2\u003e\u003cp\u003eA systematic reduction of the baseline model architecture was performed to create a computationally efficient variant. The network depth was reduced from four to two encoding-decoding blocks, and the channel dimensions were scaled down to 8, 16, and 32 in the bottleneck layer. The convolutional blocks were simplified by removing redundant layers and implementing depthwise separable convolutions where appropriate. This architectural optimization resulted in a substantial reduction of parameters while maintaining the essential spatial processing capabilities required for 3D data.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003e3.3.2 Weight Pruning\u003c/h2\u003e\u003cp\u003eA magnitude-based pruning approach was applied to eliminate redundant parameters from the trained models. The pruning threshold was empirically determined through iterative experimentation, with weights below the 0.02 magnitude threshold set to zero. This process was conducted in a structured manner to preserve the functional integrity of the feature extraction layers. The pruning mask was applied globally across all convolutional layers, with periodic retraining cycles to recover any potential accuracy loss from the pruning operation.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section3\"\u003e\u003ch2\u003e3.3.3 Quantization\u003c/h2\u003e\u003cp\u003ePost-training quantization was implemented to reduce the numerical precision of model parameters and activations. The 32-bit floating-point representations were converted to 8-bit integer format using a symmetric quantization scheme. Calibration was performed using a representative subset of the training data to determine appropriate scaling factors for each layer. The quantization process focused on the convolutional layers and linear transformations within the network, with special consideration for the skip connection operations to maintain gradient flow during inference.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e3.4 Implementation Details\u003c/h2\u003e\u003cp\u003eAll experiments were conducted using PyTorch 2.0 with CUDA acceleration where available. The training process utilized a batch size of 4 due to memory constraints of the voxel representations. Data augmentation techniques including random rotation and scaling were applied during training to improve model robustness. The optimization techniques were implemented sequentially, with each optimization step followed by a brief fine-tuning phase to stabilize performance. The entire framework was designed with modularity in mind, allowing for independent application of each optimization technique or their combined implementation as required by specific deployment scenarios.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e3.5 Evaluation Metrics\u003c/h2\u003e\u003cp\u003eThe performance of all models was evaluated using multiple metrics to provide comprehensive assessment. Model size was measured in megabytes, including all parameters and necessary metadata for deployment. Inference time was measured in milliseconds using standardized input tensors on consistent hardware configurations. Parameter efficiency was calculated as the ratio of active parameters to total model capacity. Additional metrics included memory footprint during inference and computational complexity in terms of floating-point operations required for a single forward pass. These metrics collectively provided a holistic view of model efficiency and suitability for edge deployment.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Results and Discussion","content":"\u003cp\u003eThis section presents a comprehensive evaluation of the EdgeFusion framework, comparing the performance of optimized models against the baseline across multiple metrics. The results demonstrate the effectiveness of our optimization strategies for enabling 3D diffusion models on resource-constrained devices.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1 Experimental Setup and Evaluation Metrics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll experiments were conducted on a standardized testing environment using the ModelNet40 test set comprising 2,468 3D models. Performance was evaluated using four key metrics: model size (MB), parameter count, inference time (ms), and memory footprint during operation. Inference time was measured as the average processing time for 100 consecutive forward passes using batch size 1 to simulate real-time edge deployment scenarios.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 Baseline Model Performance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe baseline 3D diffusion model established the performance reference point for subsequent comparisons. The model achieved functional 3D generation capability with a parameter count of 1,460,993 and a storage size of 5.58 MB. The average inference time measured 2.47 milliseconds per forward pass on the test hardware. While this baseline provided satisfactory generation quality, its computational demands rendered it unsuitable for deployment on typical edge devices with limited memory and processing capabilities.\u003c/p\u003e\n\u003cp\u003eFigure 1 provides a comprehensive comparison of all model variants across four key performance metrics: model size, inference time, parameter count, and speedup factor. The visualization clearly demonstrates the progressive improvements achieved through each optimization technique.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Individual Optimization Results\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe evaluation of individual optimization techniques revealed distinct performance characteristics and efficiency gains. Architectural pruning emerged as the most impactful strategy, achieving a 97.0% reduction in both parameters (from 1.46M to 43.7K) and model size (from 5.58 MB to 0.17 MB) while improving inference speed by 2.7\u0026times;, demonstrating that fundamental network redesign delivers the most substantial benefits for edge deployment. Weight pruning successfully sparsified the model by zeroing out 90.2% of parameters (1,315,561 weights) and achieved a 1.2\u0026times; speedup, though it provided limited storage reduction due to implementation constraints that maintain the original parameter structure. Quantization delivered more modest improvements with a 1.1\u0026times; speedup and theoretical 50% storage reduction potential, showing particular promise for specialized hardware environments optimized for integer operations, though its benefits in general computing scenarios were less pronounced.\u003c/p\u003e\n\u003cp\u003eTable 1 presents detailed numerical metrics for all model configurations, quantifying the exact improvements in model size, parameter count, inference time, and efficiency gains relative to the baseline.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e: Detailed metrics for quantitative analysis\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3.1 Architectural Pruning Effectiveness\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe architectural pruning strategy yielded the most significant improvements in model efficiency. By reducing network depth and channel dimensions, the optimized architecture achieved a parameter count of 43,665, representing a 97.0% reduction compared to the baseline. The model size decreased to 0.17 MB while maintaining the core 3D generation functionality. Inference time improved to 0.91 milliseconds, representing a 2.7\u0026times; speedup over the baseline. This demonstrates that careful architectural design can achieve substantial efficiency gains without compromising fundamental model capabilities.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3.2 Weight Pruning Performance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe magnitude-based weight pruning approach successfully identified and eliminated redundant parameters while preserving model functionality. The pruning process zeroed out 1,315,561 parameters, achieving 90.2% sparsity in the weight matrices. Interestingly, while the parameter count reduction was substantial, the model size remained at 5.58 MB due to the storage format requirements. The inference time improved to 2.13 milliseconds, representing a 1.2\u0026times; speedup. This result highlights that weight pruning primarily benefits computational efficiency rather than storage requirements in standard deployment scenarios.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3.3 Quantization Impact\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe 8-bit quantization implementation demonstrated moderate improvements in model efficiency. The quantized model maintained the same parameter count as the baseline but achieved a theoretical storage reduction potential of 50% (to 2.79 MB) when using integer-only arithmetic. The inference time improved to 2.27 milliseconds, representing a 1.1\u0026times; speedup. The modest performance gain suggests that quantization provides greater benefits in specialized hardware environments optimized for integer operations, which are common in edge computing platforms.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 EdgeFusion Combined Optimization\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe integrated EdgeFusion framework, combining all three optimization strategies, achieved the most compelling results for edge deployment. The final model attained a parameter count of 43,505 with a model size of 0.17 MB, representing a 97.0% reduction in both metrics compared to the baseline. The inference time reached 0.63 milliseconds, delivering a 3.9\u0026times; speedup. This comprehensive optimization resulted in a model that occupies only 170 KB of storage space while maintaining the essential 3D generation capabilities required for practical applications.\u003c/p\u003e\n\u003cp\u003eFigure 2 analyzes the optimization trade-offs, showing the relationship between model size and inference speed across different approaches, along with the individual impact of each optimization technique on model size reduction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.5 Performance Trade-off Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe evaluation reveals important trade-offs between the optimization techniques. Architectural pruning provided the most substantial benefits across all metrics, fundamentally redesigning the network for efficiency. Weight pruning offered computational improvements but limited storage benefits due to implementation constraints. Quantization showed potential for specialized hardware but provided modest gains in general computing environments. The combined EdgeFusion approach demonstrated that synergistic integration of these techniques can achieve near-optimal performance across all evaluation metrics.\u003c/p\u003e\n\u003cp\u003eTable 2 summarizes the characteristics and effectiveness of each optimization technique, providing a clear overview of how architectural pruning, weight pruning, quantization, and their combination contribute to the final EdgeFusion performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e: Optimization techniques summary and effectiveness\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003c/strong\u003e\u003cstrong\u003e4.6 Generation Quality Assessment\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eQualitative analysis of generated samples confirmed that the optimized models maintained functional 3D generation capabilities despite the substantial parameter reduction. While some detail loss was observable in complex geometries, the core shape representation remained intact for most object categories. This balance between efficiency and capability makes the EdgeFusion framework particularly suitable for applications where computational constraints outweigh the need for high-frequency detail, such as mobile AR/VR applications and real-time design tools.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.7 Discussion and Implications\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe results demonstrate that 3D diffusion models can be effectively optimized for edge deployment without sacrificing core functionality. The 97% model size reduction and 3.9\u0026times; inference speedup achieved by EdgeFusion represent significant advancements toward practical 3D AI on resource-constrained devices. These improvements enable new application scenarios in mobile computing, embedded systems, and real-time 3D content generation that were previously limited by computational requirements.\u003c/p\u003e\n\u003cp\u003eThe success of computer vision systems in agricultural applications, such as Ugwumba\u0026apos;s (2025) biomass estimation framework, reinforces our findings that well-optimized models can deliver robust performance in practical edge computing scenarios. This parallel development across domains suggests a broader trend toward making advanced AI capabilities available outside traditional high-performance computing environments\u003c/p\u003e\n\u003cp\u003eHence, the success of architectural pruning suggests that model efficiency should be considered from the initial design phase rather than as a post-training optimization. The complementary nature of the optimization techniques indicates that holistic approaches combining multiple strategies yield superior results compared to individual optimizations. This insight provides valuable guidance for future research in efficient 3D model design and deployment.\u003c/p\u003e\n\u003cp\u003eThe EdgeFusion framework establishes a foundation for bringing advanced 3D generation capabilities to edge devices, potentially enabling new applications in mobile gaming, augmented reality, and on-device content creation tools. The demonstrated efficiency gains make real-time 3D generation feasible on hardware previously considered insufficient for such tasks, expanding the accessibility and applicability of 3D AI technologies.\u003c/p\u003e"},{"header":"5. Conclusion and Future Work","content":"\u003cp\u003eThis research has successfully demonstrated that significant optimization of 3D diffusion models for edge deployment is not only possible but can be achieved with remarkable efficiency. The proposed EdgeFusion framework has addressed the critical challenge of computational intensity in 3D generative AI by systematically implementing and combining architectural pruning, weight sparsification, and quantization techniques. The results conclusively show that our approach reduces model size by 97.0%, decreases parameter count from 1.46\u0026nbsp;million to 43.5 thousand, and achieves a 3.9\u0026times; inference speedup while maintaining essential 3D generation capabilities. These achievements represent a substantial advancement toward making sophisticated 3D content generation accessible on resource-constrained devices, thereby bridging the gap between state-of-the-art generative AI and practical edge computing applications.\u003c/p\u003e\u003cp\u003eLooking forward, the success of reinforcement learning approaches in optimization tasks, such as the Deep Q-Network for task prioritization by Ugwumba \u0026amp; Jaja (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), suggests promising directions for enhancing our EdgeFusion framework. Future iterations could employ similar reinforcement learning techniques to automatically discover optimal model compression strategies that dynamically balance the trade-offs between model size, inference speed, and generation quality.\u003c/p\u003e\u003cp\u003eThe implications of this work extend across multiple domains, including mobile augmented reality, real-time design tools, and embedded 3D visualization systems. By reducing the computational barriers to 3D generative AI, EdgeFusion enables new applications where processing power, memory availability, and energy consumption are limiting factors. The framework's modular design allows for flexible adaptation to various deployment scenarios, from smartphones to specialized edge hardware.\u003c/p\u003e\u003cp\u003eFor future work, several promising directions emerge. First, investigating the application of EdgeFusion to more complex 3D representations beyond voxel grids, such as neural radiance fields or implicit neural representations, could further expand its utility. Second, exploring hardware-aware neural architecture search could automate and optimize the model design process for specific edge platforms. Third, developing dynamic optimization strategies that adapt model complexity based on available resources would enhance practical deployment in variable computing environments. Finally, extending the framework to support conditional generation and multi-modal inputs would increase its applicability to real-world scenarios requiring interactive and context-aware 3D content creation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthical Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. This research did not involve human participants, animal subjects, or any primary data collection from living entities.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests, financial or non-financial, relevant to the content of this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors received no specific funding for this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthorship Contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNnaemeka KIngsley Ugwumba: Conceptualization, Methodology, Software, Writing - Original Draft.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe author reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Declaration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analysed during this study, including the figures and source code, are available in the following GitHub repository https://github.com/KingsleyTechie/EdgeFusion3D\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBalraj. (2019). ModelNet40 [Data set]. Kaggle. https://www.kaggle.com/datasets/balraj98/modelnet40-princeton-3d-object-dataset \u003c/li\u003e\n\u003cli\u003eDavid, R., Duke, J., Jain, A., Reddi, V. J., Jeffries, N., Li, J., Kreeger, N., Nappier, I., Natraj, M., Regev, S., Rhodes, R., Wang, T., \u0026amp; Warden, P. (2021). TensorFlow Lite Micro: Embedded machine learning for TinyML. Journal of Machine Learning Research, 22(1), 1-24. https://doi.org/10.5555/3454287.3454456 \u003c/li\u003e\n\u003cli\u003eDevlin, J., Chang, M.-W., Lee, K., \u0026amp; Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 1, 4171\u0026ndash;4186. https://doi.org/10.18653/v1/N19-1423 \u003c/li\u003e\n\u003cli\u003eHan, S., Pool, J., Tran, J., \u0026amp; Dally, W. J. (2015). Learning both weights and connections for efficient neural networks. Advances in Neural Information Processing Systems, 28, 1135\u0026ndash;1143. https://doi.org/10.5555/2969239.2969364 \u003c/li\u003e\n\u003cli\u003eHe, Y., Lin, J., Liu, Z., Wang, H., Li, L.-J., \u0026amp; Han, S. (2018). AMC: AutoML for model compression and acceleration on mobile devices. Proceedings of the European Conference on Computer Vision (ECCV), 784\u0026ndash;800. https://doi.org/10.1007/978-3-030-01228-1_47 \u003c/li\u003e\n\u003cli\u003eHinton, G., Vinyals, O., \u0026amp; Dean, J. (2015). Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531. https://doi.org/10.48550/arXiv.1503.02531 \u003c/li\u003e\n\u003cli\u003eHo, J., Jain, A., \u0026amp; Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840\u0026ndash;6851. https://doi.org/10.5555/3495724.3496298 \u003c/li\u003e\n\u003cli\u003eJacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., \u0026amp; Kalenichenko, D. (2018). Quantization and training of neural networks for efficient integer-arithmetic-only inference. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2704\u0026ndash;2713. https://doi.org/10.1109/CVPR.2018.00286 \u003c/li\u003e\n\u003cli\u003eLi, H., Kadav, A., Durdanovic, I., Samet, H., \u0026amp; Graf, H. P. (2017). Pruning filters for efficient ConvNets. International Conference on Learning Representations. https://doi.org/10.48550/arXiv.1608.08710 \u003c/li\u003e\n\u003cli\u003eLuo, X., \u0026amp; Liu, Z. (2022). Diffusion-based 3D scene synthesis with neural radiance fields. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4168\u0026ndash;4178. https://doi.org/10.1109/CVPR52688.2022.00414 \u003c/li\u003e\n\u003cli\u003eRavi, D., Wong, C., Lo, B., \u0026amp; Yang, G.-Z. (2022). Deep learning for human activity recognition: A resource-efficient implementation on edge devices. IEEE Transactions on Mobile Computing, 21(4), 1234\u0026ndash;1248. https://doi.org/10.1109/TMC.2020.3034211 \u003c/li\u003e\n\u003cli\u003eUgwumba, N. K. (2025). Computer vision for pasture biomass estimation: Enabling data-driven grazing decisions through multi-modal deep learning. Research Square. https://doi.org/10.21203/rs.3.rs-8071124/v1 \u003c/li\u003e\n\u003cli\u003eUgwumba, N. K., \u0026amp; Jaja, P. S. (2025). Enhanced task prioritization system using Deep-Q-Network model. International Journal of Computer Science Engineering Techniques, 9(6), IJCSE-V9I6P15. https://doi.org/10.5281/zenodo.17636107 \u003c/li\u003e\n\u003cli\u003eWu, J., Zhang, C., Xue, T., Freeman, B., \u0026amp; Tenenbaum, J. (2016). Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. Advances in Neural Information Processing Systems, 29, 82\u0026ndash;90. https://doi.org/10.5555/3157096.3157106 \u003c/li\u003e\n\u003cli\u003eZhang, C., Patras, P., \u0026amp; Haddadi, H. (2020). Deep learning in mobile and wireless networking: A survey. IEEE Communications Surveys \u0026amp; Tutorials, 22(3), 2224\u0026ndash;2287. https://doi.org/10.1109/COMST.2019.2904897 \u003c/li\u003e\n\u003cli\u003eZhou, L., Du, Y., \u0026amp; Wu, J. (2021). 3D shape generation and completion through point-voxel diffusion. Proceedings of the IEEE/CVF International Conference on Computer Vision, 5826\u0026ndash;5835. https://doi.org/10.1109/ICCV48922.2021.00578 \u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 and 2 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Laskenta Technologies Limited","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"3D Generation, Edge Computing, AI Optimization, Model Compression, Efficient AI, Diffusion Models, Tiny Machine Learning, Mobile AI, 3D Artificial Intelligence, Lightweight Models","lastPublishedDoi":"10.21203/rs.3.rs-8168090/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8168090/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis research tackles a big problem in 3D artificial intelligence: the models that create 3D objects are usually too large and slow to run on small devices like phones or embedded systems. We present EdgeFusion, a new method that makes these 3D AI models much smaller and faster. Our approach combines several smart shrinking techniques, like removing unnecessary parts of the model and simplifying its calculations. The results are very strong. We reduced the model size by 97%, from 5.58 megabytes to just 0.17 megabytes. We also made it 3.9 times faster. This proves that it is possible to run powerful 3D generation on devices with limited resources, opening new doors for mobile and edge computing applications.\u003c/p\u003e","manuscriptTitle":"EdgeFusion: A Diffusion Framework for Real-Time 3D Generation on Resource-Constrained Devices","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-25 07:44:36","doi":"10.21203/rs.3.rs-8168090/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2ebaabe4-3c99-4688-8619-ac5d173b08b4","owner":[],"postedDate":"November 25th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":58351849,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-11-25T07:44:36+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-25 07:44:36","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8168090","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8168090","identity":"rs-8168090","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}