Figure 2. Method overview. We first apply a relighting diffusion
modelthatconvertsN imagesI ,...,I withknowncameraposes
1 N
π , ..., π , captured under extremely different illuminations, to a
1 N
set of images with the same poses, but rendered under the illumi-
nation of the reference image I (highlighted in orange). We then
1
optimize a neural radiance field to obtain a consistent 3D repre-
sentation with a novel per-image shading embedding, which can
be used to render new views of the scene from unobserved poses.