Face image generation with pose and expression control
One aspect of the disclosure is a non-transitory computer-readable storage medium including program instructions. Operations performed by execution of the program instructions include obtaining an input image that depicts a face of a subject, having an initial facial expression and an initial pose, determining a reference shape description based on the input image, determining a target shape description based on the reference shape description, a facial expression difference, and a pose difference, generating a rendered target shape image using the target shape description, and generating an output image based on the input image and the rendered target shape using an image generator, wherein the output image is a simulated image of the subject of the input image that has a final expression that is based on the initial facial expression and the facial expression difference, and a final pose that is based on the initial pose and the pose difference.
1. A non-transitory computer-readable storage medium including program instructions executable by one or more processors that, when executed, cause the one or more processors to perform operations, the operations comprising:
obtaining an input image that depicts a face of a subject, wherein the face of subject has an initial facial expression and an initial pose;
determining a reference shape description based on the input image;
determining a target shape description based on the reference shape description, a facial expression difference, and a pose difference;
generating a rendered target shape image using the target shape description; and
generating an output image based on the input image and the rendered target shape image using an image generator, wherein the output image is a simulated image of the subject of the input image that has a final expression that is based on the initial facial expression and the facial expression difference, and a final pose that is based on the initial pose and the pose difference.
2. The non-transitory computer-readable storage medium of claim 1 , wherein the image generator is a machine learning model that is trained to constrain generation of the output image based on the input image such the output image appears to depict the subject of the input image.
3. The non-transitory computer readable storage medium of claim 1 , wherein the output image is generated by a trained generator from a generative adversarial network that is trained using a discriminator that determines whether a person depicted in the output image is the subject of the input image.
4. The non-transitory computer-readable storage medium of claim 1 , wherein the rendered target shape image does not appear to depict the subject of the input image.
5. The non-transitory computer-readable storage medium of claim 1 , wherein the facial expression difference is encoded using action units from the Facial Action Coding System.
6. The non-transitory computer-readable storage medium of claim 1 , wherein the pose difference is encoded using rotation values.
7. The non-transitory computer-readable storage medium of claim 1 , the operations further comprising:
determining a reference shape description based on the input image; and
determining a target shape description based on the reference shape description, the facial expression difference, and the pose difference, wherein generating the rendered target shape image based on the input image, the facial expression difference, and the pose difference comprises generating the rendered target shape image using the target shape description.
8. The non-transitory computer-readable storage medium of claim 7 , wherein the reference shape description and the target shape description are statistical representations of face shape.
9. A system, comprising:
a memory that includes program instructions; and
a processor that is operable to execute the program instructions, wherein the program instructions, when executed by the processor, cause the processor to:
obtain an input image that depicts a face of a subject, wherein the face of the subject has an initial facial expression and an initial pose,
generate a rendered target shape image based on the input image, a facial expression difference, and a pose difference, and
generate an output image based on the input image and the rendered target shape image, wherein the output image is a simulated image of the subject of the input image that has a final expression that is based on the initial facial expression and the facial expression difference, and a final pose that is based on the initial pose and the pose difference.
10. The system of claim 9 , wherein the output image is generated by a machine learning model that is trained to constrain generation of the output image based on the input image such that the output image appears to depict the subject of the input image.
11. The system of claim 9 , wherein the program instructions further cause the processor to:
determine a reference shape description based on the input image; and
determine a target shape description based on the reference shape description, the facial expression difference, and the pose difference, wherein the program instructions cause the processor to generate the rendered target shape image based on the input image, the facial expression difference, and the pose difference by generating the rendered target shape image using the target shape description.
12. The system of claim 9 , wherein the rendered target shape image does not appear to depict the subject of the input image.
13. The system of claim 9 , wherein the facial expression difference is encoded using action units from the Facial Action Coding System.
14. The system of claim 9 , wherein the pose difference is encoded using rotation values.
15. The system of claim 11 , wherein the refrence shape description and the target shape description are statistical representations of face shape.
16. The system of claim 9 , wherein the output image is generated by a trained generator from a generative adversarial network that is trained using a discriminator that determines whether a person depicted in the output image is the subject of the input image.
17. A method, comprising:
obtaining an input image that depicts a face of a subject, wherein the face of the subject has an initial facial expression and an initial pose;
generating a rendered target shape image based on the input image, a facial expression difference, and a pose difference; and
generating an output image based on the input image and the rendered target shape image, wherein the output image is a simulated image of the subject of the input image that has a final expression that is based on the initial facial expression and the facial expression difference, and a final pose that is based on the initial pose and the pose difference.
18. The method of claim 17 , wherein the output image is generated by a machine learning model that is trained to constrain generation of the output image based on the input image such that the output image appears to depict the subject of the input image.
19. The method of claim 17 , wherein the output image is generated by a trained generator from a generative adversarial network that is trained using a discriminator that determines whether a person depicted in the output image is the subject of the input image.
20. The method of claim 17 , wherein the rendered target shape image does not appear to depict the subject of the input image.
21. The method of claim 17 , wherein the facial expression difference is encoded using action units from the Facial Action Coding System.
22. The method of claim 17 , wherein the pose difference is encoded using rotation values.
23. The method of claim 17 , further comprising:
determining a reference shape description based on the input image; and
determining a target shape description based on the reference shape description, the facial expression difference, and the pose difference, wherein generating the rendered target shape image based on the input image, the facial expression difference, and the pose difference comprises generating the rendered target shape image using the target shape description.
24. The method of claim 23 , wherein the reference shape description and the target shape description are statistical representations of face shape.