Scholarlink

We propose Latent Diffusion-Enhanced 3D Gaussian Splatting (LatentDiff-3DGS) model, an extension of 3D Gaussian Splatting (3DGS) that incorporates pretrained latent diffusion models to enhance the visual quality of novel view synthesis. By integrating a large-scale latent diffusion model into the 3DGS framework, our method captures fine structural details and improves rendering quality, addressing limitations seen in traditional 3DGS approaches. To maintain computational efficiency, we introduce a novel training strategy that applies the diffusion-related loss every k iterations, reducing the computational overhead typically associated with diffusion models. Experimental results on the Tandt [7] dataset demonstrate significant improvements in both qualitative and quantitative metrics, including PSNR, SSIM, and LPIPS, compared to the baseline. LatentDiff-3DGS provides an efficient solution for high-quality novel view synthesis with minimal computational overhead.

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

Leveraging Latent Diffusion in 3D Gaussian Splatting for Novel View Synthesis

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Bohan Li,
  • Xingyi Li,
  • Yangwen Liang,
  • Shuangquan Wang,
  • Kee-Bong Song

摘要

We propose Latent Diffusion-Enhanced 3D Gaussian Splatting (LatentDiff-3DGS) model, an extension of 3D Gaussian Splatting (3DGS) that incorporates pretrained latent diffusion models to enhance the visual quality of novel view synthesis. By integrating a large-scale latent diffusion model into the 3DGS framework, our method captures fine structural details and improves rendering quality, addressing limitations seen in traditional 3DGS approaches. To maintain computational efficiency, we introduce a novel training strategy that applies the diffusion-related loss every k iterations, reducing the computational overhead typically associated with diffusion models. Experimental results on the Tandt [7] dataset demonstrate significant improvements in both qualitative and quantitative metrics, including PSNR, SSIM, and LPIPS, compared to the baseline. LatentDiff-3DGS provides an efficient solution for high-quality novel view synthesis with minimal computational overhead.