Neural network for image style translation

1. A computer-implemented method comprising:

receiving a training input including a pair of keyframes of a training video sequence and an unpaired image of the training video sequence, the pair of keyframes representing a visual translation of an image, wherein a first keyframe in the pair of keyframes is a first version of the image in a source visual domain and a second keyframe in the pair of keyframes is a second version of the image in a target visual domain, and wherein the unpaired image is in the source visual domain;

generating, by an image translation network, a first training image from the first version of the image and a second training image from the unpaired image, wherein the first training image and the second training image are generated in the target visual domain based on the second version of the image; and

training the image translation network to translate images from the source visual domain to the target visual domain based on a calculated loss using the first training image and the second training image.

2. The computer-implemented method of claim 1 , wherein generating the first training image from the first version of the image and the second training image from the unpaired image comprises:

generating the first training image for the first version of the image by translating the first version of the image from the source visual domain to the target visual domain; and

generating the second training image for the unpaired image by translating the unpaired image from the source visual domain to the target visual domain.

3. The computer-implemented method of claim 1 , wherein the image translation network is configured to generate the second training image by:

analyzing the unpaired image to identify contents of the unpaired image;

analyzing the second version of the image from the pair of keyframes to identify a visual style of the second version of the image; and

applying the visual style of the second version of the image to the contents of the unpaired image.

4. The computer-implemented method of claim 1 , wherein training the image translation network based on the calculated loss comprises:

calculating, using a first loss function, a first loss of the first training image and the second version of the image from the pair of keyframes;

calculating, using a second loss function, a second loss of the second training image and the second version of the image from the pair of keyframes; and

determining the calculated loss from the first loss and the second loss.

5. The computer-implemented method of claim 4 , wherein calculating the first loss of the first training image and the second version of the image from the pair of keyframes comprises:

minimizing a sum of the absolute differences between the first training image and the second version of the image from the pair of keyframes.

6. The computer-implemented method of claim 4 , wherein calculating the second loss of the second training image and the second version of the image from the pair of keyframes comprises:

measuring a visual consistency between the second training image and the second version of the image from the pair of keyframes.

7. The computer-implemented method of claim 1 , further comprising:

receiving an input including a video sequence and a request to translate frames of the video sequence from the source visual domain to the target visual domain;

sending the frames of the video sequence to the trained image translation network to translate the frames of the video sequence from the source visual domain to the target visual domain; and

returning an output including a translated version of the video sequence in the target visual domain.

8. The computer-implemented method of claim 1 , further comprising:

receiving a second training input including a second pair of keyframes of a source image and a second unpaired image, the second pair of keyframes of the source image representing a visual translation of the source image, wherein a third keyframe in the second pair of keyframes is a first version of a portion of the source image in a second source visual domain and a fourth keyframe in the second pair of keyframes is a second version of the portion of the source image in a second target visual domain, and wherein the second unpaired image is in the source visual domain;

generating, by the image translation network, a third training image from the first version of the portion of the source image and a fourth training image from the second unpaired image, wherein the third training image and the fourth training image are generated in the second target visual domain based on the second version of the portion of the source image; and

training the image translation network to translate images from the second source visual domain to the second target visual domain based on a calculated loss using the third training image and the fourth training image.

9. A non-transitory computer-readable storage medium including instructions stored thereon which, when executed by at least one processor, cause the at least one processor to:

receive a training input including a pair of keyframes of a training video sequence and an unpaired image of the training video sequence, the pair of keyframes representing a visual translation of an image, wherein a first keyframe in the pair of keyframes is a first version of the image in a source visual domain and a second keyframe in the pair of keyframes is a second version of the image in a target visual domain, and wherein the unpaired image is in the source visual domain;

generating, by an image translation network, a first training image from the first version of the image and a second training image from the unpaired image, wherein the first training image and the second training image are generated in the target visual domain based on the second version of the image; and

train the image translation network to translate images from the source visual domain to the target visual domain based on a calculated loss using the first training image and the second training image.

10. The non-transitory computer-readable storage medium of claim 9 , wherein generating the first training image from the first version of the image and the second training image from the unpaired image comprises:

generating the first training image for the first version of the image by translating the first version of the image from the source visual domain to the target visual domain; and

generating the second training image for the unpaired image by translating the unpaired image from the source visual domain to the target visual domain.

11. The non-transitory computer-readable storage medium of claim 9 , wherein the image translation network is configured to generate the second training image by:

analyzing the unpaired image to identify contents of the unpaired image;

analyzing the second version of the image from the pair of keyframes to identify a visual style of the second version of the image; and

applying the visual style of the second version of the image to the contents of the unpaired image.

12. The non-transitory computer-readable storage medium of claim 9 , wherein training the image translation network based on the calculated loss comprises:

calculating, using a first loss function, a first loss of the first training image and the second version of the image from the pair of keyframes;

calculating, using a second loss function, a second loss of the second training image and the second version of the image from the pair of keyframes; and

determining the calculated loss from the first loss and the second loss.

13. The non-transitory computer-readable storage medium of claim 12 , wherein calculating the first loss of the first training image and the second version of the image from the pair of keyframes comprises:

minimizing a sum of the absolute differences between the first training image and the second version of the image from the pair of keyframes.

14. The non-transitory computer-readable storage medium of claim 12 , wherein calculating the second loss of the second training image and the second version of the image from the pair of keyframes comprises:

measuring a visual consistency between the second training image and the second version of the image from the pair of keyframes.

15. The non-transitory computer-readable storage medium of claim 9 , further comprising:

receiving an input including a video sequence and a request to translate frames of the video sequence from the source visual domain to the target visual domain;

sending the frames of the video sequence to the trained image translation network to translate the frames of the video sequence from the source visual domain to the target visual domain; and

returning an output including a translated version of the video sequence in the target visual domain.

16. The non-transitory computer-readable storage medium of claim 9 , further comprising:

receiving a second training input including a second pair of keyframes of a source image and a second unpaired image, the second pair of keyframes of the source image representing a visual translation of the source image, wherein a third keyframe in the second pair of keyframes is a first version of a portion of the source image in a second source visual domain and a fourth keyframe in the second pair of keyframes is a second version of the portion of the source image in a second target visual domain, and wherein the second unpaired image is in the source visual domain;

generating, by the image translation network, a third training image from the first version of the portion of the source image and a fourth training image from the second unpaired image, wherein the third training image and the fourth training image are generated in the second target visual domain based on the second version of the portion of the source image; and

training the image translation network to translate images from the second source visual domain to the second target visual domain based on a calculated loss using the third training image and the fourth training image.

17. A computer-implemented method comprising:

receiving, by a machine-learning backed service, an input including a video sequence and a request to translate frames of the video sequence from a source visual domain to a target visual domain;

sending the frames of the video sequence to an image translation network trained to translate the frames of the video sequence from the source visual domain to the target visual domain, the image translation network trained by generating training images in a target visual domain from a training input that includes a pair of keyframes of a training video sequence and an unpaired image of the training video sequence, the pair of keyframes representing a visual translation of an image, wherein a first keyframe in the pair of keyframes is a first version of the image in a source visual domain and a second keyframe in the pair of keyframes is a second version of the image in the target visual domain, wherein the unpaired image is in the source visual domain; and

returning an output including a translated version of the video sequence in the target visual domain.

18. The computer-implemented method of claim 17 , wherein the image translation network is trained using a training system configured to:

generate, by the image translation network, a first training image from the first version of the image and a second training image from the unpaired image, wherein the first training image and the second training image are generated in the target visual domain based on the second version of the image; and

train the image translation network to translate images from the source visual domain to the target visual domain based on a calculated loss using the first training image and the second training image.

19. The computer-implemented method of claim 18 , wherein training the image translation network based on the calculated loss comprises:

generating the first training image for the first version of the image by translating the first version of the image from the source visual domain to the target visual domain; and

generating the second training image for the unpaired image by translating the unpaired image from the source visual domain to the target visual domain.

20. The computer-implemented method of claim 18 , wherein the image translation network is configured to generate the second training image for the unpaired image by:

analyzing the unpaired image to identify contents of the unpaired image;

analyzing the second version of the image from the pair of keyframes to identify a visual style of the second version of the image; and

applying the visual style of the second version of the image to the contents of the unpaired image.