NeRF-Art: Text-Driven Neural Radiance Fields Stylization.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Neural Radiance Fields (NeRF) excel at novel view synthesis but stylization, especially text-guided, remains difficult.
• Existing methods struggle with simultaneous appearance and geometry alterations or require mesh guidance.

Purpose of the Study:

• To introduce NeRF-Art, a novel text-guided approach for stylizing pre-trained NeRF models.
• To achieve simultaneous manipulation of both appearance and geometry in NeRFs based on text prompts.

Main Methods:

• A global-local contrastive learning strategy combined with directional constraints for style control.
• Weight regularization to mitigate artifacts like cloudy effects and geometry noise during density field transformation.

Main Results:

• NeRF-Art successfully stylizes NeRF models using text prompts, altering both geometry and appearance.
• The method achieves high-quality stylization without relying on 3D meshes.
• Demonstrated effectiveness and robustness in single-view stylization quality and cross-view consistency across various styles.

Conclusions:

• NeRF-Art offers a robust and effective solution for text-guided 3D scene stylization.
• The approach overcomes limitations of previous methods by enabling simultaneous geometry and appearance changes without mesh guidance.