Photometric registration from an arbitrary geometry for augmented reality is performed using video frames of an environment captured by a camera. A surface reconstruction of the environment is generated. A pose is determined for the camera with respect to the environment, e.g., using model based tracking using the surface reconstruction. Illumination data for the environment is determined from a video frame. Estimated lighting conditions for the environment are generated based on the surface reconstruction and the illumination data. For example, the surface reconstruction may be used to compute the possible radiance transfer, which may be compressed, e.g., using spherical harmonic basis functions, and used in the lighting conditions estimation. A virtual object may then be rendered based on the lighting conditions. Differential rendering may be used with lighting solutions from the surface reconstruction of the environment and a second surface reconstruction of the environment combined with the virtual object.