Identity management in image generation

1 . A system comprising:

an image attribute estimator module implemented on a first processor in communication with a first memory, wherein the processor is to execute specific computer-executable instructions to receive a first face image depicting a face and associated with an identity and generate a set of estimated attributes of the first face image;

a person image generator module implemented on a second processor in communication with a second memory, wherein the processor is to execute specific computer-executable instructions to:

receive a clothing image depicting an article of clothing;

encode, using a clothing image encoder, the clothing image in order to extract one or more shape features comprising first details of the shape of the article of clothing;

generate, using a first U-Net encoder, one or more person features, wherein the one or more person features comprise second details of a model wearing the article of clothing;

encode, using a second VAE encoder, a person image to extract one or more clothing features from the person image, wherein the one or more clothing features comprise third details of the article of clothing; and

decode, using a first VAE decoder, the one or more person features and the one or more clothing features to generate a blurred face output, wherein the blurred face output depicts the model wearing the article of clothing and having a blurred face portion;

a face deblurring module implemented on a third processor in communication with a third memory, wherein the processor is to execute specific computer-executable instructions to:

encode, using a third VAE encoder, the blurred face output to extract one or more blurred face features;

encode, using a second U-Net encoder, the one or more blurred face features to extract one or more sharp face features conditioned on the identity; and

decode, using a second VAE decoder, the one or more sharp face features to generate a second face image depicting a sharp face conditioned on the identity; and

a final image processing module implemented on a fourth processor in communication with a fourth memory, wherein the processor is to execute specific computer-executable instructions to combine the generated second face image with the blurred face output to generate a final image depicting the identity wearing the article of clothing.

2 . The system of claim 1 , wherein the set of estimated attributes of the first face image comprises one or more of color, texture, or shape of the face.

3 . The system of claim 1 , wherein the one or more person features further comprise one or more of color, texture, or shape of one or more body parts of the model.

4 . The system of claim 1 , wherein the one or more clothing features comprise one or more of color, texture, or shape of the article of clothing.

5 . The system of claim 1 , wherein first dimensions of the second face image are adjusted to fit second dimensions of a face region of the blurred face output.

6 . A computer-implemented method comprising:

receiving a first face image depicting a face associated with an identity;

receiving a clothing image depicting an article of clothing;

estimating attributes of the face from the first face image;

generating a blurred face output image based on encodings of the clothing image and the estimated attributes of the face, wherein the blurred face output image depicts a model with a blurred face wearing the article of clothing;

extracting a face region around the blurred face from the blurred face output image using semantic segmentation;

deblurring the face region into a sharp real face conditioned on the identity; and

combining the sharp real face into the blurred face output image.

7 . The computer-implemented method of claim 6 , wherein encodings of the clothing image comprise one or more person features comprising one or more clothing features comprising one or more of color, texture, or shape of the article of clothing.

8 . The computer-implemented method of claim 6 , wherein attributes of the face include one or more of color, texture, or shape of the face.

9 . The computer-implemented method of claim 6 , wherein deblurring the face region comprises:

encoding the blurred face output image to extract one or more blurred face features using a first VAE encoder;

encoding the first face image to extract one or more sharp face features using a U-Net encoder; and

decoding the one or more sharp face features using a VAE decoder.

10 . The computer-implemented method of claim 6 , wherein the face region is deblurred based on the estimated attributes of the face.

11 . The computer-implemented method of claim 6 , further comprising:

receiving a second face image comprising a second face;

accessing a set of face images associated with the identity;

determining a third face image from the set of face images with a similar pose to the second face image;

extracting a homography of the second face image and the third face image;

applying perspective transformation to the homography to deform the third face image; and

integrating the second face image onto the deformed third face image.

12 . The computer-implemented method of claim 6 , wherein the sharp real face conditioned on the identity is based on encoding a set of face images including the first face image using an encoder.

13 . The computer-implemented method of claim 12 , wherein the clothing image is encoded using a first clothing image encoder to extract one or more clothing shape features and using a second clothing image encoder to extract one or more clothing features.

14 . The computer-implemented method of claim 13 , wherein generating the blurred face output image comprises:

generating, using a U-Net, one or more person features based on the one or more clothing shape features and the estimated attributes of the face; and

decoding the one or more person features and the one or more clothing features to generate the blurred face output using a VAE decoder.

15 . A non-transitory computer-readable medium storing computer-executable instructions that, when executed by a processor of a computing device, cause the computing device to at least:

estimate attributes of a face associated with an identity from a first face image depicting the face;

generate a blurred face output image based on the estimated attributes of the face and encodings of a clothing image depicting an article of clothing, wherein the blurred face output image depicts a model with a blurred face wearing the article of clothing;

extract a face region around the blurred face from the blurred face output image using semantic segmentation;

deblur the face region into a sharp real face conditioned on the identity; and

integrate the sharp real face into the blurred face output image.

16 . The non-transitory computer-readable medium of claim 15 , wherein the first face image and the clothing image are received as input.

17 . The non-transitory computer-readable medium of claim 15 , wherein encodings of the clothing image comprise one or more person features comprising one or more of color, texture, or shape of one or more body parts and one or more clothing features comprising one or more of color, texture, or shape of the article of clothing.

18 . The non-transitory computer-readable medium of claim 15 , wherein attributes of the face include one or more of color, texture, or shape of the face.

19 . The non-transitory computer-readable medium of claim 15 , wherein deblurring the face region comprises:

encode the blurred face output to extract one or more blurred face features using a first VAE encoder;

encode the first face image to extract one or more sharp face features using a second VAE encoder; and

decode the one or more blurred face features and the one or more sharp face features using a VAE decoder.

20 . The non-transitory computer-readable medium of claim 15 , wherein the face region is deblurred based on the estimated attributes of the face.

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

receive a second face image comprising a second face and hair;

access a set of face images associated with the identity;

determine a third face image from the set of face images with a similar pose to the second face image;

extract a homography of the second face image and the third face image;

apply perspective transformation to the homography to deform the third face image; and

integrate the hair of the second face image onto the deformed third face image.

22 . The non-transitory computer-readable medium of claim 15 , wherein the sharp real face conditioned on the identity is based on encoding a set of face images including the first face image using an encoder.

23 . The non-transitory computer-readable medium of claim 22 , wherein the clothing image is encoded using a first clothing image encoder to extract one or more clothing shape features and using a second clothing image encoder to extract one or more clothing features.

24 . The non-transitory computer-readable medium of claim 23 , wherein generating the blurred face output image comprises:

generate, using a U-Net, one or more person features based on the one or more clothing shape features and the estimated attributes of the face; and

decode the one or more person features and the one or more clothing features to generate the blurred face output using a VAE decoder.