PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
May 20, 2026Journal of Computer-Aided Design & Computer Graphics0 citationsOpen Access

3DGS Generation and Color Editing with Rich Text

QWQiling WuKXKun Xu

Key Points

  • The aim is to enhance the efficiency and consistency of semantic editing in 3D Gaussian models, particularly for color corrections.
  • Developed a rich text-guided semantic editing method for 3D Gaussian models.
  • Generated attention maps and semantic segmentation for precise color editing.
  • Implemented a regional diffusion denoising process to improve the color transition.
  • Compared to two reference methods, the proposed approach achieved more precise editing with reduced color leakage.
  • Editing time decreased from 10 minutes to just 25 seconds per instance.
  • Results demonstrated a more natural appearance in the edited areas.

Abstract

三维高斯溅射凭借其渲染效率与质量,已经成为生成式三维模型的主流表示。然而,现有的三维高斯的语义编辑方法效率低、一致性差,而基于二维扩散的编辑方法容易产生三维不一致、颜色渗漏等问题。因此,提出一种基于富文本引导的三维高斯语义编辑方法,重点是色彩编辑任务。首先,使用现有的文本生成三维高斯模型得到待编辑高斯与跨视角注意力图;其次,利用注意力图提取三维一致的语义分割,建立富文本词组与编辑区域的关联;最后,设计基于区域的扩散降噪过程,结合语义分割进行噪声混合,逐步引导编辑区域的色彩向目标色彩逼近。在多条提示词上的实验结果表明,所提方法和2种参考方法相比,编辑区域更为精准、颜色渗漏更少,且编辑结果更自然;和基于优化的方法相比,单次编辑时间从10 min缩短至25 s。

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Wu et al. (2025) studied this question.

synapsesocial.com/papers/6a0d50f3f03e14405aa9d212https://doi.org/10.3724/sp.j.1089.2025-00261
Ask AI
Helpful
Bookmark
Share
View Full Paper