Face anonymization in digital images
Face anonymization techniques are described that overcome conventional challenges to generate an anonymized face. In one example, a digital object editing system is configured to generate an anonymized face based on a target face and a reference face. As part of this, the digital object editing system employs an encoder as part of machine learning to extract a target encoding of the target face image and a reference encoding of the reference face. The digital object editing system then generates a mixed encoding from the target and reference encodings. The mixed encoding is employed by a machine-learning model of the digital object editing system to generate a mixed face. An object replacement module is used by the digital object editing system to replace the target face in the target digital image with the mixed face.
1. A method comprising:
receiving, by a processing device, a target digital image having a target face;
obtaining, by the processing device, a reference digital image including a reference face as a result of a digital image search performed based at least in part on the target face by:
computing a similarity score between the target face and the reference face; and
comparing a pose of the target face to a pose of the reference face;
generating, by the processing device, a target encoding of the target face and a reference encoding of the reference face using a machine-learning model;
generating, by the processing device, a mixed encoding by combining a portion of the target encoding and a portion of the reference encoding that represents features of the target face mixed with features of the reference face;
generating, by the processing device, a mixed face using a machine-learning model from the mixed encoding; and
forming, by the processing device, an edited target digital image by replacing the target face with the mixed face.
2. The method as described in claim 1 , further comprising generating a search query that includes the target face and receiving a search result that includes the reference digital image responsive to an image search performed using the search query.
3. The method as described in claim 2 , wherein the search result includes a plurality of candidate digital images and further comprising receiving a user input selecting at least one of the plurality of candidate digital images as the reference digital image.
4. The method as described in claim 2 , wherein the image search is based at least in part on color, pose, and feature descriptors.
5. The method as described in claim 1 , wherein the target encoding and the reference encoding are formed as latent vectors and the generating the mixed encoding is performed using linear interpolation.
6. The method as described in claim 1 , wherein the generating the mixed face using the machine-learning model is implemented as a decoder as part of a neural network to convert latent vectors of the mixed encoding into pixels forming the mixed face.
7. The method as described in claim 1 , further comprising receiving a user input specifying an amount of the target encoding or the reference encoding to be used in the generating of the mixed encoding.
8. The method as described in claim 7 , further comprising outputting a user control in a user interface usable to specify the amount of the target encoding or the reference encoding to be used in the generating of the mixed encoding and wherein the receiving of the user input is performed responsive to user interaction with the user control.
9. The method as described in claim 1 , further comprising adding additional facial skin details to the mixed face.
10. In a digital medium digital object mixing environment, a system comprising:
an encoder module implemented at least partially in hardware of a processing device to generate a target encoding of a target digital object in a target digital image and a reference encoding of a reference digital object in a reference digital image by:
computing a similarity score between the target digital object and the reference digital object; and
comparing a pose of the target digital object to a pose of the reference digital object;
an object mix module implemented at least partially in hardware of the processing device to generate a mixed encoding by combining a portion of the reference encoding and a portion of the target encoding that represents features of the target digital object mixed with features of the reference digital object;
a generator module implemented at least partially in hardware of the processing device to generate a mixed digital object from the mixed encoding using a machine-learning model; and
an object replacement module implemented at least partially in hardware of the processing device to generate an edited digital image by replacing the target digital object with the mixed digital object in the target digital image.
11. The system as described in claim 10 , further comprising a search query module implemented at least partially in hardware of the processing device to generate a search query including the target digital object and receive a search result as a result of an image search that includes the reference digital image.
12. The system as described in claim 11 , wherein the image search is based at least in part on color, pose, and feature descriptors.
13. The system as described in claim 11 , further comprising a search query module configured to output the search result including a plurality of candidate digital images and receive a user input selecting at least one of the plurality of candidate digital images as the reference digital image.
14. The system as described in claim 10 , wherein the target encoding and the reference encoding are formed as latent vectors and the object mix module is configured to generate the mixed encoding using linear interpolation.
15. The system as described in claim 10 , wherein the machine-learning model of the generator module is implemented as a decoder as part of a neural network to convert latent vectors of the mixed encoding into pixels forming the mixed digital object.
16. The system as described in claim 10 , wherein the target digital object and the reference digital object are human faces.
17. In a digital medium face anonymization environment, a system comprising:
means for generating a search query based on a target face in a target digital image;
means for obtaining a reference digital image including a reference face as a result of a digital image search performed based on the search query by:
computing a similarity score between the target face and the reference face; and
comparing a pose of the target face to a pose of the reference face;
means for generating a target encoding of the target face and a reference encoding of the reference face using an encoder of a machine-learning model;
means for generating a mixed encoding by combining a portion of the target encoding and a portion of the reference encoding that represents features of the target face mixed with features of the reference face;
means for generating an anonymized face using a decoder of a machine-learning model from the mixed encoding that anonymizes the target face; and
means for editing the target digital image by replacing the target face with the anonymized face.
18. The system as described in claim 17 , wherein the image search is based at least in part on color, pose, and feature descriptors.
19. The system as described in claim 17 , wherein the target encoding and the reference encoding are formed as latent vectors and the means for generating the mixed encoding uses linear interpolation.
20. The system as described in claim 17 , further comprising means for receiving a user input specifying an amount of the target encoding or the reference encoding to be used in generating the mixed encoding.