NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections

Martin-Brualla et al. extend NeRF (Mildenhall et al., 2020) to the "in-the-wild" setting: tourist photo collections of famous landmarks where images differ in time of day, weather, exposure, and frequently contain transient occluders such as people and vehicles. Vanilla NeRF assumes a static scene with consistent appearance, and produces severe artifacts under either kind of variation.

NeRF-W introduces two extensions. A per-image appearance embedding is concatenated to the radiance MLP input, letting the network model photometric variation across captures as a low-dimensional latent code while keeping geometry shared. A separate transient head, also conditioned on a per-image embedding, predicts a transient density and color along with an uncertainty field; the loss is downweighted in pixels where the transient head asserts high uncertainty, automatically masking out photo-bombers, tourists, and scaffolding without per-image annotation.

The two components together let NeRF-W reconstruct landmarks such as Trevi Fountain, the Brandenburg Gate, and Sacre Coeur from in-the-wild Flickr image collections, producing temporally consistent free-viewpoint renderings with controllable illumination conditions. Compared to NRW and Neural Rerendering in the Wild, NeRF-W substantially improves PSNR and perceptual metrics on the Phototourism benchmark.

The appearance-embedding trick became standard in subsequent radiance- field methods that needed to absorb capture-time variation, including Block-NeRF for city-scale scenes and Mega-NeRF for aerial captures. The transient-uncertainty idea reappears in robust NeRF training, where it is used to ignore other forms of inconsistent observation. Within this corpus, NeRF-W is the first work to push NeRF beyond the controlled laboratory capture into raw consumer imagery.