View on GitHub

Research Review Notes

Summaries of academic research papers

Fader Networks: Manipulating Images by Sliding Attributes


Idea

The authors attempt to disentangle facial features from images and re-generate images after tuning (fader knobs) certain continuous-valued attributes of the image like age, expression, gender etc. This is an encoder-decoder architecture.

The major difference touted compared to existing methods is that adversarial training is used to learn the latent space, as opposed to the decoder output, thus, helping the latent space become invariant to the attributes (conditioning labels).

Method

Observations