# [InterFaceGAN: Interpreting the DisentangledFace Representation Learned by GANs](https://arxiv.org/pdf/2005.09635.pdf)
The authors analyse the latent space of StyleGAN and propose a manipulation method using incersion to latent code and projection on required direction.

### Latent Space Semantics :
**Single Semantic:**
- Interpolation from one latent code to another defines a hyperplane which serves as a separation boundary for a binary semantic.(gender)
- The distance of sample z from hyperplane is defined as n^Tz where n is normal unit vector of the hyperplane.
- The properties of hyperplane are:

**Multiple Semantics :**
- Similar to single semantic scenario another distance function is used using a multivariate normal distribution.

### Latent Space Manipulation:
- For single attribute manipulation z is changed as z+ αn, in the direction of attribute.
- For conditional manipulation we try to make the separation bounaries (N), and N^TN as orthogonal, if they arent we project them and then manipulate the attributes.
- The model works on fixed GAN ,so for real image manipulation we infer the best latent code using GAN inversion.

### Experiments:
- The model is tested on PCGAN and StyleGAN for different metrics.
- The separability of latent space is clearly found as the distance from separation boundary increases.
- For semantic manipulation if the distance from separation boundary is too high the samples dont match the original example.
- The unnatural artifacts could we fixed by training linear SVM to find suitable separation boundary and manipulate the latent code.
- The disentanglement analysis showed subspace projection reduces the entanglement by suitable amount and also the bias of certaion attributes in few datasets.