Super-NeRF: View-Consistent Detail Generation for NeRF Super-Resolution

Area of Science:

• Computer Vision
• 3D Scene Reconstruction
• Neural Rendering

Background:

• Neural Radiance Fields (NeRF) excel at synthesizing novel views of 3D scenes with high fidelity.
• Current NeRF methods primarily focus on high-resolution output from high-resolution input, neglecting low-resolution input scenarios.
• NeRF super-resolution is an underexplored but crucial area for generating detailed 3D scenes from limited-resolution data.

Purpose of the Study:

• To introduce Super-NeRF, a novel method for generating high-resolution NeRF representations from low-resolution input images.
• To address the gap in NeRF super-resolution by enabling high-fidelity detail generation from low-resolution inputs.
• To develop a method that ensures view consistency in the generated high-resolution details.

Main Methods:

• Proposed Super-NeRF, a multi-view consistency-controlling super-resolution module for NeRF.
• Introduced optimizable latent codes per view to control high-resolution 2D image generation while maintaining view consistency.
• Synergistically optimized latent codes with the Super-NeRF representation, leveraging inherent NeRF view consistency constraints.

Main Results:

• Demonstrated the effectiveness of Super-NeRF on synthetic, real-world, and AI-generated NeRF datasets.
• Achieved state-of-the-art performance in NeRF super-resolution, particularly in generating high-resolution details.
• Showcased superior cross-view consistency in the super-resolved NeRF outputs.

Conclusions:

• Super-NeRF effectively generates high-resolution NeRF from low-resolution images, significantly advancing NeRF super-resolution.
• The method demonstrates strong potential for applications requiring detailed 3D scene reconstruction from limited-resolution data.
• Super-NeRF sets a new benchmark for high-resolution detail generation and view consistency in NeRF super-resolution.