# Block-NeRF: Scalable Large Scene Neural View Synthesis

**Authors:** Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, Henrik Kretzschmar  
**arXiv:** [2202.05263](https://arxiv.org/abs/2202.05263)  
**Date:** 2022-02-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.05263v1/

## Key points

- Block-NeRF is a variant of Neural Radiance Fields (NeRF) that can represent large-scale environments.
- It decomposes scenes into individually trained NeRFs, allowing for rendering to scale to arbitrarily large environments.
- Per-block updates provide flexibility in modifying specific areas without affecting the entire scene.
- Architectural changes are introduced to make NeRF robust to variations in appearance and lighting conditions.
- Appearance embeddings, learned pose refinement, and controllable exposure are added to each individual NeRF.
- Alignment of appearance between adjacent NeRFs ensures smooth transitions and creates a cohesive representation of the entire scene.
- The authors built a grid of Block-NeRFs using 2.8 million images, creating the largest neural scene representation to date.
- This representation can render an entire neighborhood of San Francisco.

## Summary

In their paper titled "Block-NeRF: Scalable Large Scene Neural View Synthesis," Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar present a variant of Neural Radiance Fields (NeRF) called Block-NeRF that can effectively represent large-scale environments. They address the challenge of scaling NeRF to render city-scale scenes spanning multiple blocks by decomposing the scene into individually trained NeRFs. The authors demonstrate that this decomposition approach is crucial as it decouples rendering time from scene size and enables rendering to scale to arbitrarily large environments. It also allows for per-block updates of the environment, providing flexibility in modifying specific areas without affecting the entire scene. To make NeRF robust to data captured over months under different environmental conditions, several architectural changes are introduced. Appearance embeddings, learned pose refinement, and controllable exposure are added to each individual NeRF. These additions enhance the model's ability to handle variations in appearance and lighting conditions. Additionally, the authors propose a procedure for aligning appearance between adjacent NeRFs so that they can be seamlessly combined. This alignment ensures smooth transitions between blocks and creates a cohesive representation of the entire scene. To evaluate their approach, the authors build a grid of Block-NeRFs using 2.8 million images, creating the largest neural scene representation to date. This representation is capable of rendering an entire neighborhood of San Francisco. Overall, Block-NeRF offers a scalable solution for neural view synthesis in large-scale environments by decomposing scenes into individually trained NeRFs and incorporating architectural enhancements for robustness and seamless combination of adjacent blocks.