Scholarlink

<p>The generation of three-dimensional (3D) objects from textual descriptions holds significant promise across various domains, including healthcare, architecture, and virtual reality. However, existing Text-to-3D generation methods often face challenges related to data scarcity, computational inefficiency, and the fidelity of generated models. Here, we introduce SwiftCraft3D, a novel framework designed to address these challenges through a two-stage paradigm. In the first stage, we employ Semantic-Enhanced Multi-View Visual Prompts Generation, where input text is semantically enriched using GPT-4 and used to generate multi-view visual prompts via a fine-tuned diffusion model. This ensures that the visual prompts are geometrically consistent and semantically aligned with the input text. The second stage, Text-Infused Sparse-View 3D Reconstruction, integrates these visual prompts with the original text to produce high-fidelity 3D models using a multi-modal fusion strategy and the Triplane Decoder. Extensive experiments demonstrate that SwiftCraft3D not only achieves superior geometric consistency and detail in the generated 3D models but also significantly reduces computational overhead compared to state-of-the-art methods, achieving results in approximately 25&#xa0;s. Our code is available at <a href="https://github.com/OpenMICG/SwiftCraft3D">https://github.com/OpenMICG/SwiftCraft3D</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

SwiftCraft3D: semantic-enhanced multi-view prompting for efficient and high-fidelity text-to-3D generation

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Zeyi Dong,
  • Ting Yu

摘要

The generation of three-dimensional (3D) objects from textual descriptions holds significant promise across various domains, including healthcare, architecture, and virtual reality. However, existing Text-to-3D generation methods often face challenges related to data scarcity, computational inefficiency, and the fidelity of generated models. Here, we introduce SwiftCraft3D, a novel framework designed to address these challenges through a two-stage paradigm. In the first stage, we employ Semantic-Enhanced Multi-View Visual Prompts Generation, where input text is semantically enriched using GPT-4 and used to generate multi-view visual prompts via a fine-tuned diffusion model. This ensures that the visual prompts are geometrically consistent and semantically aligned with the input text. The second stage, Text-Infused Sparse-View 3D Reconstruction, integrates these visual prompts with the original text to produce high-fidelity 3D models using a multi-modal fusion strategy and the Triplane Decoder. Extensive experiments demonstrate that SwiftCraft3D not only achieves superior geometric consistency and detail in the generated 3D models but also significantly reduces computational overhead compared to state-of-the-art methods, achieving results in approximately 25 s. Our code is available at https://github.com/OpenMICG/SwiftCraft3D.