Image composites using a generative neural network
The present disclosure relates to an image composite system that employs a generative adversarial network to generate realistic composite images. For example, in one or more embodiments, the image composite system trains a geometric prediction neural network using an adversarial discrimination neural network to learn warp parameters that provide correct geometric alignment of foreground objects with respect to a background image. Once trained, the determined warp parameters provide realistic geometric corrections to foreground objects such that the warped foreground objects appear to blend into background images naturally when composited together.
1. In in a digital environment for creating or editing digital content, a computer-implemented method for determining geometrically correct warp parameters, the computer-implemented method comprising:
utilizing a geometric prediction neural network to determine initial warp parameters based on a geometric perspective of a background image;
generating, by at least one processor, an initial composite image utilizing the geometric prediction neural network by combining the background image and a foreground object transformed with the initial warp parameters;
determining updated warp parameters for the foreground object utilizing the geometric prediction neural network; and
generating an updated composite image utilizing the geometric prediction neural network by combining the background image and the foreground object transformed with the updated warp parameters.
2. The computer-implemented method of claim 1 , further comprising utilizing the geometric prediction neural network to determine the updated warp parameters for the foreground object based on the initial warp parameters, the foreground object, and the geometric perspective of a background image.
3. The computer-implemented method of claim 1 , further comprising determining the initial warp parameters for the foreground object in low-dimensional space as initial low-dimensional warp parameters.
4. The computer-implemented method of claim 3 , further comprising:
transforming a high-resolution version of the foreground object utilizing the initial low-dimensional warp parameters; and
generating a high-resolution initial composite image by combining the high-resolution version of the foreground object transformed with the initial low-dimensional warp parameters and a high-resolution version of the background image.
5. The computer-implemented method of claim 1 , further comprising:
utilizing the geometric prediction neural network to generate an object mask for the foreground object transformed with the initial warp parameters; and
generating the initial composite image by combining the background image, the foreground object transformed with the initial warp parameters, and the object mask.
6. The computer-implemented method of claim 1 , wherein:
the geometric prediction neural network is a spatial transformer convolutional neural network; and
the geometric prediction neural network comprises a generator network that generates composite images by combining images of foreground objects to background images based on determined warp parameters.
7. The computer-implemented method of claim 6 , further comprising:
utilizing an adversarial discrimination neural network that distinguishes realistic images from non-realistic images by classifying digital images as real images or fake images; and
wherein determining the updated warp parameters for the foreground object utilizing the geometric prediction neural network is based on the adversarial discrimination neural network classifying the initial composite image as a fake image.
8. The computer-implemented method of claim 7 , wherein:
the adversarial discrimination neural network comprises a discriminator network that determines if the generated composite images resemble realistic images; and
the geometric prediction neural network and the adversarial discrimination neural network form a spatial transformer-generative adversarial network (ST-GAN).
9. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor, cause a computer system to:
generate an initial composite image generated based on a foreground object, initial warp parameters, and a background image; and
generate an updated composite image utilizing a geometric prediction neural network by:
determining updated warp parameters for the foreground object based on the initial warp parameters and a geometric perspective of the background image;
spatially transform the foreground object by applying the updated warp parameters to the foreground object; and
combining the spatially transformed foreground object with the background image to generate the updated composite image.
10. The non-transitory computer-readable medium of claim 9 , further comprising instructions that, when executed by the at least one processor, cause the computer system to identify the foreground object by:
identify a foreground image that comprises the foreground object;
generate an object mask that isolates the foreground object within the foreground image; and
apply the object mask to the foreground image to identify the foreground object.
11. The non-transitory computer-readable medium of claim 9 , further comprising instructions that, when executed by the at least one processor, cause the computer system to utilize the geometric prediction neural network to generate the initial warp parameters based on a geometric perspective of the background image.
12. The non-transitory computer-readable medium of claim 11 , further comprising instructions that, when executed by the at least one processor, cause the computer system to generate the initial composite image utilizing the geometric prediction neural network by combining the background image and the foreground object after the foreground object has been spatially transformed based on the initial warp parameters.
13. The non-transitory computer-readable medium of claim 9 , further comprising instructions that, when executed by the at least one processor, cause the computer system to:
receive a fake image classification for the initial composite image from an adversarial discrimination neural network that distinguishes realistic images from non-realistic images; and
based on the fake image classification, generate the updated composite image utilizing the geometric prediction neural network.
14. The non-transitory computer-readable medium of claim 9 , further comprising instructions that, when executed by the at least one processor, cause the computer system to determine the initial warp parameters for the foreground object in low-dimensional space as initial low-dimensional warp parameters.
15. The non-transitory computer-readable medium of claim 14 , further comprising instructions that, when executed by the at least one processor, cause the computer system to:
transform a high-resolution version of the foreground object utilizing the initial low-dimensional warp parameters; and
generate a high-resolution initial composite image by combining the high-resolution version of the foreground object transformed with the initial low-dimensional warp parameters and a high-resolution version of the background image.
16. A system for training a neural network to create realistic geometric corrections for image composting using adversarial training comprising:
a memory comprising:
a background image;
a foreground object; and
a geometric prediction neural network;
at least one processor; and
at least one processor configured to cause the system to:
determine initial warp parameters for the foreground object utilizing the geometric prediction neural network and a geometric perspective of the background image;
initially transform the foreground object based on the initial warp parameters;
generate an initial composite image utilizing the geometric prediction neural network by combining the background image and the initially transformed foreground object;
determine updated warp parameters for the foreground object utilizing the geometric prediction neural network, the initial warp parameters, and the geometric perspective of the background image;
generate an updated composite image utilizing the geometric prediction neural network by combining the background image with the foreground object after the foreground object has been transformed with the updated warp parameters; and
provide the updated composite image to a client device.
17. The system of claim 16 , wherein the at least one processor is further configured to cause the system to iteratively generate the updated warp parameters at the geometric prediction neural network to improve spatial geometry of the foreground object has been transformed with the updated warp parameters when composited with the background image.
18. The system of claim 16 , wherein the at least one processor is further configured to cause the system to constrain the initial warp parameters to reside within a trust region of the background image by utilizing penalties within an objective function of the geometric prediction neural network.
19. The system of claim 16 , wherein the at least one processor is further configured to cause the system to:
provide the initial composite image to an adversarial discrimination neural network to determine if the initial composite image resembles a realistic image comprising a naturally placed foreground object and a background; and
based on the initial composite image not resembling a realistic image, determine the warp parameters for the foreground object utilizing the geometric prediction neural network.
20. The system of claim 16 , wherein the at least one processor is further configured to cause the system to initially position the foreground object within the background image based on user input before determining warp parameters for the foreground object.