The paper introduces a novel approach for improving NeRF inpainting performance by leveraging a latent diffusion model and addressing optimization challenges through a masked adversarial training scheme. The proposed method aims to enhance the completion of missing regions in NeRF-rendered scenes by incorporating a diffusion prior and optimizing the inpainting process using adversarial training with masked regions. The key components of the approach include masked adversarial loss, feature matching loss, per-scene customization, and the utilization of pixel-level and perceptual loss functions.

The implementation details involve training the NeRF model using a dataset consisting of scenes with objects to be removed, along with corresponding inpainting masks. The training process optimizes the NeRF model based on the training views with objects present, while evaluation is performed on test views where the objects are physically removed. The evaluation metrics include LPIPS for perceptual difference, M-LPIPS for inpainting performance, and FID/KID for distributional similarity between rendered and ground-truth images. Additionally, a custom evaluation metric, C-FID/C-KID, is used for scenes without physically removed objects.