Producing higher-quality samples of natural images

1. A method comprising, by one or more computing devices:

accessing a plurality of generative adversarial networks (GANs) that are each applied to a particular level k of a Laplacian pyramid, wherein each GAN comprises a generative model G k and a discriminative model D k , wherein, at each level k:

the generative model G, takes as input a noise vector z k and outputs a generated image {tilde over (h)} k ; and

the discriminative model D k takes as input either the generated image {tilde over (h)} k or a real image h k drawn from a set of training data, and outputs a probability that the input was the real image h k , wherein the set of training data comprises:

a first set of color images of a plurality of objects belonging to a plurality of different classes, wherein the images in the first set are tightly cropped around the respective objects; and

a second set of color images of a plurality of natural scene types, wherein the images in the second set are not tightly cropped around any object:

generating a sample image Ĩ k from the generated images {tilde over (h)} k , wherein the sample image is based at least in part on the probabilities outputted by each of the discriminative models D k at each level k of the Laplacian pyramid and the generated images {tilde over (h)} k ; and

providing the sample image Ĩ k for display on a display screen of a client device of a user of a communications network.

2. The method of claim 1 , wherein generating the sample image Ĩ k from the generated images {tilde over (h)} k further comprises:

applying an upsampling operator u(.) to at least one of the generated images {tilde over (h)} k ; and

combining the upsampled image {tilde over (h)} k with an image I k+1 generated by a generative model G k+1 , wherein the image I k+1 was generated at level k+1 in the Laplacian pyramid.

3. The method of claim 2 , wherein the sample image Ĩ k is generated using the equation Ĩ k =u(Ĩ k+1 )+{tilde over (h)} k .

4. The method of claim 1 , wherein the generated images {tilde over (h)} k are generated in a coarse-to-fine fashion.

5. The method of claim 1 , wherein each level k in the Laplacian pyramid corresponds to a generated image {tilde over (h)} k that comprises a particular number of pixels, wherein as k increases, the number of pixels in {tilde over (h)} k decreases.

6. The method of claim 1 , wherein the generative model G k and the discriminative model D k each take an additional vector of information as input.

7. The method of claim 6 , wherein the additional vector of information is a generated image l k created by a conditional GAN model.

8. One or more computer-readable non-transitory storage media embodying software that is operable when executed to:

access a plurality of generative adversarial networks (GANs) that are each applied to a particular level k of a Laplacian pyramid, wherein each GAN comprises a generative model G and a discriminative model D k , wherein, at each level k:

the generative model G k takes as input a noise vector z k and outputs a generated image {tilde over (h)} k ; and

the discriminative model D k , takes as input either the generated image {tilde over (h)} k or a real image h k drawn from a set of training data, and outputs a probability that the input was the real image h k , wherein the set of training data comprises:

a first set of color images of a plurality of objects belonging to a plurality of different classes, wherein the images in the first set are tightly cropped around the respective objects; and

a second set of color images of a plurality of natural scene types, wherein the images in the second set are not tightly cropped around any object;

generate a sample image Ĩ k from the generated images {tilde over (h)} k , wherein the sample image is based at least in part on the probabilities outputted by each of the discriminative models D k at each level k of the Laplacian pyramid and the generated images {tilde over (h)} k ; and

provide the sample image Ĩ k for display on a display screen of a client device of a user of a communications network.

9. The media of claim 8 , wherein generating the sample image Ĩ k from the generated images {tilde over (h)} k , further comprises:

applying an upsampling operator u(.) to at least one of the generated images {tilde over (h)} k ; and

combining the upsampled image {tilde over (h)} k with an image I k+1 generated by a generative model G k+1 , wherein the image I k+1 was generated at level k+1 in the Laplacian pyramid.

10. The media of claim 9 , wherein the sample image Ĩ k is generated using the equation Ĩ k =u(Ĩ k+1 )+{tilde over (h)} k .

11. The media of claim 8 , wherein the generated images {tilde over (h)} k are generated in a coarse-to-fine fashion.

12. The media of claim 8 , wherein each level k in the Laplacian pyramid corresponds to a generated image {tilde over (h)} k that comprises a particular number of pixels, wherein as k increases, the number of pixels in {tilde over (h)} k decreases.

13. The media of claim 8 , wherein the generative model G k and the discriminative model D k each take an additional vector of information as input.

14. The media of claim 13 , wherein the additional vector of information is a generated image l k created by a conditional GAN model.

15. A system comprising:

one or more processors; and

one or more computer-readable non-transitory storage media coupled to one or more of the processors and comprising instructions operable when executed by one or more of the processors to cause the system to:

access a plurality of generative adversarial networks (GANs) that are each applied to a particular level k of a Laplacian pyramid, wherein each GAN comprises a generative model G k and a discriminative model D k , wherein, at each level k:

the generative model G k takes as input a noise vector z k and outputs a generated image {tilde over (h)} k ; and

the discriminative model D k takes as input either the generated image {tilde over (h)} k or a real image h k drawn from a set of training data, and outputs a probability that the input was the real image h k , wherein the set of training data comprises:

a first set of color images of a plurality of objects belonging to a plurality of different classes, wherein the images in the first set are tightly cropped around the respective objects; and

a second set of color images of a plurality of natural scene types, wherein the images in the second set are not tightly cropped around any object:

generate a sample image Ĩ k from the generated images {tilde over (h)} k , wherein the sample image is based at least in part on the probabilities outputted by each of the discriminative models D k at each level k of the Laplacian pyramid and the generated images {tilde over (h)} k ; and

provide the sample image Ĩ k for display on a display screen of a client device of a user of a communications network.

16. The system of claim 15 , wherein generating the sample image Ĩ k from the generated images {tilde over (h)} k further comprises:

applying an upsampling operator u(.) to at least one of the generated images {tilde over (h)} k ; and

combining the upsampled image {tilde over (h)} k with an image I k+1 generated by a generative model G k+1 , wherein the image I k+1 was generated at level k+1 in the Laplacian pyramid.

17. The system of claim 16 , wherein the sample image Ĩ k is generated using the equation Ĩ k =u(Ĩ k+1 )+{tilde over (h)} k .

18. The system of claim 15 , wherein the generated images {tilde over (h)} k are generated in a coarse-to-fine fashion.

19. The system of claim 15 , wherein each level k in the Laplacian pyramid corresponds to a generated image {tilde over (h)} k that comprises a particular number of pixels, wherein as k increases, the number of pixels in {tilde over (h)} k decreases.

20. The system of claim 15 , wherein the generative model G k and the discriminative model D k each take an additional vector of information as input.