GaussEdit: Adaptive 3D Scene Editing With Text and Image Prompts

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• 3D scene editing is complex, often requiring specialized skills.
• Existing methods struggle with maintaining global coherence during local edits.
• The Janus problem (view-dependent inconsistencies) remains a challenge.

Purpose of the Study:

• To introduce GaussEdit, a novel framework for adaptive 3D scene editing.
• To enable precise scene modifications guided by text and image prompts.
• To improve the accuracy, visual fidelity, and efficiency of 3D scene editing.

Main Methods:

• Utilizes 3D Gaussian Splatting for scene representation.
• Employs a three-stage editing process: initialization, adaptive global-local optimization, and texture enhancement.
• Incorporates category-guided regularization to address the Janus problem.

Main Results:

• GaussEdit demonstrates superior performance over existing methods in editing accuracy and visual fidelity.
• The framework achieves faster processing speeds.
• Successfully integrates user-specified concepts into 3D scenes with realistic results.

Conclusions:

• GaussEdit provides a powerful and user-driven approach to 3D scene editing.
• The adaptive optimization and texture synthesis ensure high-quality, prompt-aligned edits.
• Offers significant improvements over traditional 3D scene editing techniques.