Inverse Rendering

Inverse rendering works backward from images or other measurements to estimate the scene properties and physical processes that produced them, enabling reconstruction, relighting, and editing. Approaches include differentiable rasterization and ray tracing with meshes, neural fields, or Gaussian splats, as well as learned priors and feed-forward models that predict scene properties directly. Researchers evaluate these methods on synthetic and real captures using geometry and material accuracy, novel-view synthesis, and relighting under changed illumination. A central challenge is that different combinations of geometry, materials, and lighting can explain the same observations, especially with sparse views, so methods must balance physical consistency and generalization against speed and incomplete data.