**GFP-GAN** is a generative adversarial network for blind face restoration that leverages a generative facial prior (GFP). This Generative Facial Prior (GFP) is incorporated into the face restoration process via channel-split spatial feature transform layers, which allow for a good balance between realness and fidelity. As a whole, the GFP-GAN consists of a degradation removal module ([U-Net](https://paperswithcode.com/method/u-net)) and a pretrained face  [StyleGAN](https://paperswithcode.com/method/stylegan) as a facial prior. They are bridged by a latent code mapping and several Channel-Split [Spatial Feature Transform](https://paperswithcode.com/method/spatial-feature-transform) (CS-SFT) layers. During training, 1) intermediate restoration losses are employed to remove complex degradation, 2) Facial component loss with discriminators is used to enhance facial details, and 3) identity preserving loss is used to retain face identity.

**MoGA-A** is a convolutional neural network optimized for mobile latency and discovered via Mobile GPU-Aware (MoGA) [neural architecture search](https://paperswithcode.com/method/neural-architecture-search). The basic building block is MBConvs (inverted residual blocks) from [MobileNetV2](https://paperswithcode.com/method/mobilenetv2). Squeeze-and-excitation layers are also experimented with.

MoGA-A

MoGA: Searching Beyond MobileNetV3

GFP-GAN

Towards Real-World Blind Face Restoration with Generative Facial Prior

**Adafactor** is a stochastic optimization method based on [Adam](https://paperswithcode.com/method/adam) that reduces memory usage while retaining the empirical benefits of adaptivity. This is achieved through maintaining a factored representation of the squared gradient accumulator across training steps. Specifically, by tracking moving averages of the row and column sums of the squared gradients for matrix-valued variables, we are able to reconstruct a low-rank approximation of the exponentially smoothed accumulator at each training step that is optimal with respect to the generalized Kullback-Leibler divergence. For an $n \times m$ matrix, this reduces the memory requirements from $O(n m)$ to $O(n + m)$. 

Instead of defining the optimization algorithm in terms of absolute step sizes {$\alpha_t$}$\_{t=1}^T$, the authors define the optimization algorithm in terms of relative step sizes {$\rho_t$}$\_{t=1}^T$, which get multiplied by the scale of the parameters. The scale of a parameter vector or matrix is defined as the root-mean-square of its components, lower-bounded by a small constant $\epsilon_2$.  The reason for this lower bound is to allow zero-initialized parameters to escape 0. 

Proposed hyperparameters are: $\epsilon\_{1} = 10^{-30}$, $\epsilon\_{2} = 10^{-3}$, $d=1$, $p\_{t} = \min\left(10^{-2}, \frac{1}{\sqrt{t}}\right)$, $\hat{\beta}\_{2\_{t}} = 1 - t^{-0.8}$.

Source	Towards Real-World Blind Face Restoration with Generative Facial Prior
Year	2000
Data Source	CC BY-SA - https://paperswithcode.com