GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields

Michael Niemeyer, Andreas Geiger


The paper presents a generative neural feature fields based approach for representing scenes as compositional generative neural feature fields for the disentanglement of objects from the background as well as individual object shapes and appearances while learning from unstructured and unposed image collections. The model learns a disentangled representation of individual objects in a scene by using separate latent vectors for each object and allows rotation, translation, and change in the camera pose of objects.




Horizontal, vertical translations and rotations performed using shape and appearance latent vector interpolations show the disentanglement of the background and the cars.

Our Two Cents


Project Page: