Image to 3D Generation Platform

Abstract—This paper presents a complete web-based platform that transforms single images or text prompts into downloadable textured 3D meshes in GLB format. The system follows a production-ready workflow including user authentication, dash- board interface, input preprocessing with background removal and segmentation, structured 3D generation using sparse voxel representations, mesh refinement, real-time browser preview, and object storage integration. A novel selective INT8 weight-only quantization technique is applied specifically to transformer flow modules, achieving approximately 33% model size reduction and 50% lower GPU memory footprint while maintaining drop-in compatibility with the original architecture. The platform also incorporates asynchronous job orchestration, GPU worker pools, metadata tracking, and comprehensive monitoring. Experimental evaluation on representative single-object datasets demonstrates superior inference speed compared to optimization-based meth- ods like DreamFusion while preserving competitive visual quality. The proposed system demonstrates practical deployment feasi- bility for applications in AR commerce, gaming asset generation, educational visualization, and digital twin creation.