Generating images using neural networks

1. A system comprising:

one or more computers, and

one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations for generating an output image comprising a plurality of pixels arranged in a two-dimensional map, each pixel having a respective value for each of one or more channels, the operations comprising:

receiving a neural network input that includes a high-level description of desired content of the output image;

generating, using one or more initial neural network layers and conditioned on the high-level description, a respective portion of an alternative representation of the neural network input for each of the one or more channels for each pixel in the plurality of pixels included in the output image; and

for each of the one or more channels for each pixel in the plurality of pixels included in the output image:

processing the respective portion of the alternative representation using one or more output neural network layers to generate a respective score distribution over a discrete set of possible values for the channel; and

selecting, using the respective score distribution, the respective value for the channel.

2. The system of claim 1 , wherein the high-level description of the desired content of the output image is represented as a latent vector that specifies a class of object that should appear in the output image.

3. The system of claim 2 , wherein the latent vector is a one-hot encoded vector.

4. The system of claim 1 , wherein the neural network input includes the high-level description of the desired content of the output image and any already generated pixels for the output image.

5. The system of claim 1 , wherein the one or more initial neural network layers comprise one or more convolutional neural network layers.

6. The system of claim 5 , wherein the one or more initial neural network layers further comprise one or more gated non-linear activation layers that are each configured to receive as input a combination of a corresponding output of the convolutional neural network layer and an input generated from the high-level description and to process the combination to generate a gated non-linear activation layer output by computing an element-wise multiplication between an activation function output and a gate vector that is generated by applying an element-wise non-linearity to the corresponding output of the convolutional neural network layer.

7. The system of claim 1 , wherein the one or more channels are ordered, wherein the one or more output neural network layers comprise a respective output layer corresponding to each of the one or more channels, and wherein each of the output layers is configured to, for each pixel of the output image:

generate the respective score distribution over the discrete set of possible values for the channel corresponding to the output layer conditioned on (i) values for pixels for pixels before the pixel in the sequence and (ii) values for the pixel for any channels before the channel corresponding to the output layer in the order of channels.

8. The system of claim 1 , wherein, for each channel of each pixel, each of the output neural network layers is configured to receive a portion of the alternative representation corresponding to the channel generated based on (i) values for pixels before the pixel in the sequence and (ii) values for the pixel for any channels before the channel in the order of channels.

9. The system of claim 8 , wherein the system is configured to apply a mask to outputs of neural network layers in the one or more initial neural network layers such that the portion of the alternative representation corresponding to the channel is only generated based on pixels before the pixel in the sequence and channel data for the pixel for channels before the channel in the order of channels.

10. One or more non-transitory computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for generating an output image comprising a plurality of pixels arranged in a two-dimensional map, each pixel having a respective value for each of one or more channels, the operations comprising:

receiving a neural network input that includes a high-level description of desired content of the output image;

generating, using one or more initial neural network layers and conditioned on the high-level description, a respective portion of an alternative representation of the neural network input for each of the one or more channels for each pixel in the plurality of pixels included in the output image; and

for each of the one or more channels for each pixel in the plurality of pixels included in the output image:

processing the respective portion of the alternative representation using one or more output neural network layers to generate a respective score distribution over a discrete set of possible values for the channel; and

selecting, using the respective score distribution, the respective value for the channel.

11. The computer-readable storage media of claim 10 , wherein the high-level description of the desired content of the output image is represented as a latent vector that specifies a class of object that should appear in the output image.

12. The computer-readable storage media of claim 11 , wherein the latent vector is a one-hot encoded vector.

13. The computer-readable storage media of claim 10 , wherein the neural network input includes the high-level description of the desired content of the output image and any already generated pixels for the output image.

14. The computer-readable storage media of claim 10 , wherein the one or more initial neural network layers comprise one or more convolutional neural network layers.

15. The computer-readable storage media of claim 14 , wherein the one or more initial neural network layers further comprise one or more gated non-linear activation layers that are each configured to receive as input a combination of a corresponding output of the convolutional neural network layer and an input generated from the high-level description and to process the combination to generate a gated non-linear activation layer output by computing an element-wise multiplication between an activation function output and a gate vector that is generated by applying an element-wise non-linearity to the corresponding output of the convolutional neural network layer.

16. The computer-readable storage media of claim 10 , wherein the one or more channels are ordered, wherein the one or more output neural network layers comprise a respective output layer corresponding to each of the one or more channels, and wherein each of the output layers is configured to, for each pixel of the output image:

generate the respective score distribution over the discrete set of possible values for the channel corresponding to the output layer conditioned on (i) values for pixels for pixels before the pixel in the sequence and (ii) values for the pixel for any channels before the channel corresponding to the output layer in the order of channels.

17. The computer-readable storage media of claim 10 , wherein, for each channel of each pixel, each of the output neural network layers is configured to receive a portion of the alternative representation corresponding to the channel generated based on (i) values for pixels before the pixel in the sequence and (ii) values for the pixel for any channels before the channel in the order of channels.

18. The computer-readable storage media of claim 17 , wherein the system is configured to apply a mask to outputs of neural network layers in the one or more initial neural network layers such that the portion of the alternative representation corresponding to the channel is only generated based on pixels before the pixel in the sequence and channel data for the pixel for channels before the channel in the order of channels.

19. A computer-implemented method of generating an output image comprising a plurality of pixels arranged in a two-dimensional map, each pixel having a respective value for each of one or more channels, and the method comprising:

receiving a neural network input that includes a high-level description of desired content of the output image;

generating, using one or more initial neural network layers and conditioned on the high-level description, a respective portion of an alternative representation of the neural network input for each of the one or more channels for each pixel in the plurality of pixels included in the output image; and

for each of the one or more channels for each pixel in the plurality of pixels included in the output image:

processing the respective portion of the alternative representation using one or more output neural network layers to generate a respective score distribution over a discrete set of possible values for the channel; and

selecting, using the respective score distribution, the respective value for the channel.

20. The computer-implemented method of claim 19 , wherein the high-level description of the desired content of the output image is represented as a latent vector that specifies a class of object that should appear in the output image.