Scholarlink

Artificial Intelligence-based Content Generation (AIGC) and Extended Reality (XR) have gained prominence for delivering immersive experiences. Recent deep learning models can reconstruct textured 3D meshes from a single image, enabling scalable content creation. However, deploying these models in real-time XR settings remains challenging due to trade-offs between quality, performance, and responsiveness. This study evaluates how three single-image-to-3D generative models CRM, Unique3D, and InstantMesh perform in immersive contexts, considering both visual fidelity and computational efficiency. We assess their outputs across four object categories (Animals, Objects, Humanoids, and Places), using quantitative metrics grouped into geometric accuracy (IoU, Chamfer Distance, Hausdorff Distance, F-Score), texture fidelity (PSNR, SSIM), perceptual realism (LPIPS, Clip Similarity, FID, CMDM, NR-3DQA), VR performance (FPS, GPU time, GPU utilization, Stale frames) and 3D model generation time and polygon count. Results show that InstantMesh achieves the best VR performance (up to 76 FPS) with low GPU usage and fast generation time (<30 s). CRM offers balanced quality and speed, while Unique3D delivers the best perceptual realism but with higher computational demands and slower generation (~60 s). This work contributes to the design of XR applications by establishing a balance between visual detail and computational efficiency, helping VR developers integrate AIGC-driven 3D objects into immersive scenarios without compromising user experience.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Analysis Between Different 3D Object Generation Models. Adaptation to Immersive Technologies

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sergio Cleger Tamayo,
  • Geovana Amorim Abensur,
  • Agustin Alejandro Ortiz Diaz,
  • Delrick Nunes de Oliveira,
  • Osvaldo Vitalino dos Santos Junior,
  • Gilberto Rufino de Oliveira Neto

摘要

Artificial Intelligence-based Content Generation (AIGC) and Extended Reality (XR) have gained prominence for delivering immersive experiences. Recent deep learning models can reconstruct textured 3D meshes from a single image, enabling scalable content creation. However, deploying these models in real-time XR settings remains challenging due to trade-offs between quality, performance, and responsiveness. This study evaluates how three single-image-to-3D generative models CRM, Unique3D, and InstantMesh perform in immersive contexts, considering both visual fidelity and computational efficiency. We assess their outputs across four object categories (Animals, Objects, Humanoids, and Places), using quantitative metrics grouped into geometric accuracy (IoU, Chamfer Distance, Hausdorff Distance, F-Score), texture fidelity (PSNR, SSIM), perceptual realism (LPIPS, Clip Similarity, FID, CMDM, NR-3DQA), VR performance (FPS, GPU time, GPU utilization, Stale frames) and 3D model generation time and polygon count. Results show that InstantMesh achieves the best VR performance (up to 76 FPS) with low GPU usage and fast generation time (<30 s). CRM offers balanced quality and speed, while Unique3D delivers the best perceptual realism but with higher computational demands and slower generation (~60 s). This work contributes to the design of XR applications by establishing a balance between visual detail and computational efficiency, helping VR developers integrate AIGC-driven 3D objects into immersive scenarios without compromising user experience.