Scholarlink

<p>The rapid development of metaverse applications has intensified the demand for efficient and user-friendly 3D content creation tools. Although sketch-based 3D modeling is a promising approach, existing methods face significant challenges, particularly in handling sparse training data and abstract sketch-to-shape pairings. To address these limitations, we propose <b>EffiSketch3D</b>, a novel framework that leverages pre-trained models for data-efficient, high-fidelity 3D generation from single-view sketches. Our approach features a two-stage architecture: (1) a data-efficient encoder-decoder for coarse shape reconstruction and (2) an octree-based diffusion refiner for detail enhancement. A key innovation is the integration of CLIP and Uni3D as discriminators operating across both 2D and 3D domains to ensure semantic consistency between the input sketch and the output shape. Furthermore, a progressive training strategy significantly reduces data requirements. Extensive experiments on the ShapeNet-Synthetic and ShapeNet-Sketch datasets demonstrate state-of-the-art performance, achieving a 56.42 mean Voxel IoU (a 6.1% improvement) and a 44.98 Chamfer Distance (a 5.9% reduction) compared to previous methods. Our work presents a robust and accessible solution for 3D content creation, well-suited for metaverse applications. The source code is available at <a href="https://github.com/77waiwai-Demon/EffiSketch3D">https://github.com/77waiwai-Demon/EffiSketch3D</a>.</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

EffiSketch3D: data-efficient 3D generation from sketches via dual-domain pre-trained priors

  • 期刊论文

  • 发表日期：

    2025年12月6日

查看全文
  • Bin Xu,
  • Xinmin Cheng,
  • Huihui Tan,
  • Yidong Han,
  • Jiahao Zhang

摘要

The rapid development of metaverse applications has intensified the demand for efficient and user-friendly 3D content creation tools. Although sketch-based 3D modeling is a promising approach, existing methods face significant challenges, particularly in handling sparse training data and abstract sketch-to-shape pairings. To address these limitations, we propose EffiSketch3D, a novel framework that leverages pre-trained models for data-efficient, high-fidelity 3D generation from single-view sketches. Our approach features a two-stage architecture: (1) a data-efficient encoder-decoder for coarse shape reconstruction and (2) an octree-based diffusion refiner for detail enhancement. A key innovation is the integration of CLIP and Uni3D as discriminators operating across both 2D and 3D domains to ensure semantic consistency between the input sketch and the output shape. Furthermore, a progressive training strategy significantly reduces data requirements. Extensive experiments on the ShapeNet-Synthetic and ShapeNet-Sketch datasets demonstrate state-of-the-art performance, achieving a 56.42 mean Voxel IoU (a 6.1% improvement) and a 44.98 Chamfer Distance (a 5.9% reduction) compared to previous methods. Our work presents a robust and accessible solution for 3D content creation, well-suited for metaverse applications. The source code is available at https://github.com/77waiwai-Demon/EffiSketch3D.