Chris Donahue, Zachary C. Lipton, Akshay Balsubramani, Julian McAuley

This is an interactive demo for the SD-BEGAN model from Semantically Decomposing the Latent Spaces of Generative Adversarial Networks (arXiv, pdf, code). This demo runs a neural network in your browser; it may take some time to render.

SD-GANs learn a latent representation that disentangles identity (ID) from the contingent factors of an observation (Obs). They can generate images that appear to depict the same person but vary in lighting, pose, expression, etc. Furthermore, they can generate images that depict distinct identities in the same contingent circumstances.


The left matrix depicts all combinations between four distinct identities (rows) and four distinct observations (columns). The right matrix depicts a linear interpolation of both the identity and observation vectors for a pair of images selected from the left matrix.

Press the buttons on the rows/columns in the left matrix to randomize identities/observations.

Select two images in the left matrix to see their interpolation in the right.

Obs A Obs B