Style-based architecture for generative neural networks
A style-based generative network architecture enables scale-specific control of synthesized output data, such as images. During training, the style-based generative neural network (generator neural network) includes a mapping network and a synthesis network. During prediction, the mapping network may be omitted, replicated, or evaluated several times. The synthesis network may be used to generate highly varied, high-quality output data with a wide variety of attributes. For example, when used to generate images of people's faces, the attributes that may vary are age, ethnicity, camera viewpoint, pose, face shape, eyeglasses, colors (eyes, hair, etc.), hair style, lighting, background, etc. Depending on the task, generated output data may include images, audio, video, three-dimensional (3D) objects, text, etc.
1. A computer-implemented method, comprising:
processing a latent code defined in an input space, by a mapping neural network, to produce an intermediate latent code defined in an intermediate latent space;
converting the intermediate latent code into a first style signal;
applying the first style signal at a first layer of a synthesis neural network to modify first intermediate data according to the first style signal to produce modified first intermediate data;
processing the modified first intermediate data to produce second intermediate data;
applying a second style signal at a second layer of the synthesis neural network to modify the second intermediate data according to the second style signal to produce modified second intermediate data; and
processing the modified second intermediate data to produce output data including content corresponding to the second intermediate data.
2. The computer-implemented method of claim 1 , wherein the first intermediate data is a constant.
3. The computer-implemented method of claim 1 , further comprising converting the intermediate latent code into the second style signal.
4. The computer-implemented method of claim 1 , further comprising:
processing a second latent code defined in the input space by the mapping neural network to produce a second intermediate latent code defined in the intermediate latent space;
converting the second intermediate latent code into the second style signal.
5. The computer-implemented method of claim 1 , wherein converting the first intermediate latent code comprises combining the first intermediate latent code with additional first intermediate latent codes produced for additional latent codes.
6. The computer-implemented method of claim 1 , wherein converting the first intermediate latent code comprises computing a weighted sum of the first intermediate latent code and additional first intermediate latent codes produced for additional latent codes.
7. The computer-implemented method of claim 1 , wherein the first layer of the synthesis neural network is an adaptive instance normalization layer.
8. The computer-implemented method of claim 1 , wherein converting the first intermediate latent code comprises performing an affine transformation on the first intermediate latent code to produce the first style signal.
9. The computer-implemented method of claim 1 , wherein a dimension of the second style signal is a multiple of a number of feature maps in the second intermediate data.
10. The computer-implemented method of claim 1 , wherein multiple variations of the output data are produced for the first intermediate data in response to the synthesis neural network receiving multiple first style signals.
11. The computer-implemented method of claim 10 , wherein a first attribute varies across the multiple variations of the output data and a second attribute is unchanged across the multiple variations of the output data.
12. The computer-implemented method of claim 1 , wherein the mapping neural network and the synthesis neural network are trained in an unsupervised manner.
13. The computer-implemented method of claim 1 , further comprising:
injecting first spatial noise into the first intermediate data; and
injecting second spatial noise into the second intermediate data.
14. The computer-implemented method of claim 13 , wherein components of the first spatial noise and components of the second spatial noise each comprise independent random variables.
15. The computer-implemented method of claim 13 , wherein components of the first spatial noise and components of the second spatial noise are each drawn from a Gaussian distribution.
16. The computer-implemented method of claim 13 , wherein the first spatial noise is multiplied by a first constant and summed with a first feature map of the first intermediate data, and by a second constant and summed with a second feature map of the first intermediate data.
17. The computer-implemented method of claim 1 , wherein the output data comprises image data.
18. The computer-implemented method of claim 1 , wherein the output data comprises audio data.
19. A system, comprising:
a mapping neural network that is configured to:
process a latent code defined in an input space to produce an intermediate latent code defined in an intermediate latent space; and
convert the intermediate latent code into a first style signal; and
a synthesis neural network that is configured to:
apply the first style signal at a first layer to modify first intermediate data according to the first style signal to produce modified first intermediate data;
process the modified first intermediate data to produce second intermediate data;
apply a second style signal at a second layer of the synthesis neural network to modify the second intermediate data according to the second style signal to produce modified second intermediate data; and
process the modified second intermediate data to produce output data including content corresponding to the second intermediate data.
20. A computer-implemented method, comprising:
applying a first set of spatial noise at a first layer of a synthesis neural network to generate modified first intermediate data comprising content corresponding to first intermediate data that is modified based on the first set of spatial noise;
processing the modified first intermediate data to produce second intermediate data;
applying a second set of spatial noise at a second layer of the synthesis neural network to generate modified second intermediate data comprising content corresponding to the second intermediate data that is modified based on the second set of spatial noise; and
processing the modified second intermediate data to produce output data comprising content corresponding to the modified second intermediate data.