Abstract
The emergence of Neural Radiance Fields (NeRF) has promoted the development of synthesized high-fidelity views of the intricate real world. However, it is still a very demanding task to repaint the content in NeRF. In this paper, we propose a novel framework that can take RGB images as input and alter the 3D content in neural scenes. Our work leverages existing diffusion models to guide changes in the designated 3D content. Specifically, we semantically select the object we want to modify first, and a pre-trained diffusion model will guide the NeRF model to generate new 3D objects, which can improve the editability, diversity, and application range of NeRF. Experiment results show that our algorithm is effective for editing 3D objects in NeRF under different text prompts, including editing appearance, shape, etc. We validate our method on real-world datasets and synthetic-world datasets for these editing tasks. See \url{https://repaintnerf.github.io} for a better view into our edited results.
Edited Pipeline
Overview of RePaint-NeRF. We present an editing method in NeRF. In the first stage, we additionally optimize a feature field along with the color module and density module to extract the content mask by using text or patch. In another way of speaking, we separate the part we want to change for a generation. Then, we use the mask and text prompt to generate the new content guided by the pre-trained diffusion model and CLIP model. After optimization of the generation, we can finally repaint a pre-trained NeRF model with view consistency and scene integrity.
BibTeX
@inproceedings{RePaint-NeRF,
title={{RePaint-NeRF}: NeRF Editting via Semantic Masks and Diffusion Models},
author={Zhou, Xingchen and He, Ying and Yu, F Richard and Li, Jianqiang and Li, You},
year={2023},
booktitle={IJCAI},
}