Start with the problem
Given several images of a scene, how can we generate an image from a viewpoint that was never photographed? This is a central question in novel-view synthesis. It relates to reconstructing an interactive physical world, but it is not the same task.
Two representative works take different approaches: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis, presented at ECCV 2020, and 3D Gaussian Splatting for Real-Time Radiance Field Rendering, presented at SIGGRAPH 2023.
NeRF: a scene as a continuous function
NeRF uses a neural network to map spatial positions and viewing directions to volume density and color. Volume rendering converts these predictions into images. Rather than merely storing photographs, the method learns a continuous scene representation that can be queried from different viewpoints.
Follow three mechanisms while reading: coordinate encoding, sampling along camera rays, and the supervision provided by comparing rendered and training images.
3D Gaussian Splatting: an explicit alternative
3D Gaussian Splatting represents a scene as a collection of 3D Gaussians, optimizing their positions, shapes, opacities and appearance. A visibility-aware splatting renderer generates images. The original work targets high-quality real-time novel-view synthesis.
This makes the implicit–explicit comparison concrete. How does the representation affect optimization, memory requirements and rendering? Does comparable image quality imply comparable geometry?
Keep three questions separate
- Image quality: how closely synthesized images match reference images.
- Geometric accuracy: whether the representation recovers reliable 3D structure.
- Physical consistency: whether the scene supports trustworthy dynamics and interaction.
Good novel-view synthesis does not automatically establish the latter two. Check hardware, resolution, datasets and training settings before interpreting performance figures. This guide makes no additional quantitative performance claims.
Original sources
- Mildenhall et al. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV 2020. Project, paper and code.
- Kerbl et al. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. SIGGRAPH 2023. Project, paper and code.
This page is a guide to public papers, not research produced by this website or a claim of collaboration with their authors.