Text-driven light-field content editing for three-dimensional light-field display based on Gaussian splatting.

Area of Science:

• Computer Graphics
• Image Processing
• Human-Computer Interaction

Background:

• Three-dimensional (3D) light-field displays offer natural stereoscopic vision but face challenges in content creation due to high production thresholds and limited user-friendly editing tools.
• Efficient and convenient generation of diverse 3D light-field content remains a significant hurdle for widespread adoption.

Purpose of the Study:

• To present a novel text-driven method for editing 3D light-field content tailored for 3D light-field displays.
• To overcome the limitations of existing methods by enabling intuitive and efficient content generation based on textual descriptions.

Main Methods:

• A text-driven editing approach utilizing Gaussian splatting for 3D light-field content generation.
• Perception fusion to propagate editing effects from key views to others, minimizing iterative editing.
• A refinement module with attention-based latent feature alignment for multi-view consistency.
• Efficient light-field coding and a generation control training strategy for direct 3D image synthesis.

Main Results:

• The proposed method rapidly generates high-quality 3D light-field content that accurately reflects text editing instructions.
• Demonstrated stability and reliability in editing results through multi-channel independent updates and average normalization.
• Achieved efficient generation of 3D light-field images directly from an edited Gaussian scene space.

Conclusions:

• The developed text-driven Gaussian splatting method significantly simplifies and accelerates the creation of 3D light-field content.
• The approach offers enhanced control and flexibility through optional user interaction extensions, catering to various content needs.
• This work paves the way for more accessible and user-friendly 3D light-field content creation.