# [Realistic Face Reconstruction from Deep Embeddings](https://openreview.net/pdf?id=-WsBmzWwPee)
The authors discuss the problems with face embedding reconstruction and proposes a method that can overcome SOTA face recognition system.

### Method
The authors propose a greedy random optimisation algorithm that searchs for matching embedding of Stylegan space to FaceNet embedding.
To improve the optimisation they use annealing.
Since random search for vector is expensive and time taking certain pairs are pre generated to improve speed.
The experiments are done on FFHQ dataset.
On comparison with Zhmoginov et Al the proposed method has better quality but the previous picks up finer details which is atrributed to the white box access.