# Bilateral Guided Radiance Field Processing

**Authors:** Yuehao Wang, Chaoyi Wang, Bingchen Gong, Tianfan Xue  
**arXiv:** [2406.00448](https://arxiv.org/abs/2406.00448)  
**Date:** 2024-06-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.00448v2/

## Key points

- Neural Radiance Fields (NeRF) revolutionized novel view synthesis by leveraging multi-view consistency for unparalleled performance
- Bilateral Guided Radiance Field Processing is a new approach to address issues while maintaining visual quality
- The method disentangles ISP enhancement effects during NeRF training and reapplies them in a finishing stage based on user preferences
- Bilateral grids are used as a generalized representation of ISP processing, optimizing per-view 3D bilateral grids with radiance fields for each input view
- A low-rank 4D bilateral grid is learned to enable user-adjustable 3D finishing and lift photo enhancements for the entire 3D scene from a single view edit
- This approach removes floaters and enhances visual quality by seamlessly incorporating user retouching into synthesized views
- Bilateral Guided Radiance Field Processing significantly improves novel view synthesis results and provides access to source code and data for further exploration

## Summary

Neural Radiance Fields (NeRF) have revolutionized the field of novel view synthesis by leveraging multi-view consistency to achieve unparalleled performance. To address this issue while maintaining visual quality, a new approach called Bilateral Guided Radiance Field Processing has been proposed. This innovative method disentangles the enhancement effects of ISP during the NeRF training stage and reapplies them to the reconstructed radiance fields at a later finishing stage based on user preferences. The key concept behind this technique is the use of bilateral grids as a generalized representation of ISP processing. By optimizing per-view 3D bilateral grids with radiance fields for each input view, the effects of camera pipelines can be approximated effectively. Moreover, to enable user-adjustable 3D finishing and lift photo enhancements to the entire 3D scene from a single view edit, a low-rank 4D bilateral grid is learned. This comprehensive approach not only removes floaters but also enhances visual quality by incorporating user retouching seamlessly into the synthesized views. The authors demonstrate that Bilateral Guided Radiance Field Processing significantly improves novel view synthesis results and provides access to their source code and data for further exploration. This research was presented at SIGGRAPH (ACM TOG) in 2024 and more information can be found on their project page at https://bilarfpro.github.io.