Learning to Factorize and Relight a City

ECCV 2020

Rotating the Sun

Modeling sun azimuth is important to our factorization for the following reasons:

  • Sun azimuth is mathematically intuitive concept, its value ranges from [0, 2π] and it encodes an orientation of the world.
  • Despite it's simplicity, changing azimuth has a complex effect on a scene's appearance.

When learning our disentangled factors, our model exploits the simplicity of the underlying representation. But when we train it to decode relit shading images, the generator must learn the rich interactions between azimuth and geometry. We show various panoramic images and the effect of manually rotating our azimuth (which is succinctly given to the generator as an angle).

We show scenes where the sun moves with the sliders below.


New York City, USA


Paris, France


San Francisco, USA


London, United Kingdom


Additional Results

We present additional generated results. Despite training on scenes from New York City, we can relight scenes from entirely new cities and even continents. Other cities coming soon!