A Precomputed Atmosphere Differentiable Renderer for Estimating Outdoor Illumination

Estimating outdoor illumination from images requires modeling the complex phenomena of atmospheric scattering. Existing methods either depend on known scene geometry or multi-view images, limiting their applicability. This work presents a differentiable renderer tailored for atmospheric scattering to recover outdoor illumination from a single image. We first construct a physics-based atmospheric scattering model to establish the analytical mapping from the sun's zenith angle to sky appearances. We propose a pre-computation scheme to cache intermediate atmospheric transmittance, scattered radiance, and ground-reflected radiance for efficient gradient-based optimization. Their gradients of the sun's zenith angle can also be precomputed and stored in derivative tables. The desired gradients can be quickly retrieved at render time to guide the optimization. Experiments on sky image rendering and illumination estimation demonstrate that our method can recover plausible outdoor illumination from a single photograph. Without the need for scene geometry or multi-view inputs, the proposed differentiable rendering framework provides an efficient way to estimate outdoor illumination.