Scholarlink

<p>To address the issues of high acquisition cost and low efficiency of 3D models, a single-view 3D model generation method based on a diffusion model and Gaussian Splatting is proposed. This method inputs a single image and supplements side views based on the diffusion model. On this basis, a 3D Gaussian Splatting (3DGS) reconstruction method based on a random view supervision strategy and a local point cloud adaptive completion mechanism is introduced to improve the reconstruction quality. The reconstructed model is rendered off-axis to generate dense viewpoint images, which are then encoded to achieve the generation of a 3D model from a single view. Experimental results show that the improved 3DGS outperforms the traditional 3DGS algorithm in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Learned Perceptual Image Similarity (LPIPS).</p>

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

ST-gaussian: improved 3D content generation based on stable diffusion and gaussian splatting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Mingtong Liu,
  • Olzhas N. Turar,
  • Mengcai Ye

摘要

To address the issues of high acquisition cost and low efficiency of 3D models, a single-view 3D model generation method based on a diffusion model and Gaussian Splatting is proposed. This method inputs a single image and supplements side views based on the diffusion model. On this basis, a 3D Gaussian Splatting (3DGS) reconstruction method based on a random view supervision strategy and a local point cloud adaptive completion mechanism is introduced to improve the reconstruction quality. The reconstructed model is rendered off-axis to generate dense viewpoint images, which are then encoded to achieve the generation of a 3D model from a single view. Experimental results show that the improved 3DGS outperforms the traditional 3DGS algorithm in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Learned Perceptual Image Similarity (LPIPS).