Choose two faces and combine their styles with AI (generative neural nets)
Additional Charges
The watermark can be removed and the generated image can be exported with an in app purchase. Each generated image requires its own in app purchase.
Here are a few ways this app can be used:
1) Hair color
Set the "input" photo to the person whose hair color you want to change. Choose a photo of a person with the hair color you want for the "style" photo. Since hair color is a relatively shallow style, styles can be mixed at shallow depths here. Mixing deeper styles will transfer deeper properties from the style photo such as face structure, gender, and head pose.
2) Baby generator
Choose photos of both parents and mix their styles. If mixing deep styles here, the gender styles will mix. There is a gender slider to adjust for this. If only mixing shallow styles, the gender will be that of the "input" face.
3) Hair transplant
This requires deeper style mixing than hair color. Set the input photo to the person who will receive the hair, and the style photo to someone who will donate the hair. Try to find the shallowest depth that transfers the hair style. This seems to be the 2nd to last and 3rd to last layers in most cases.
4) Cartoons
Some cartoon faces will be picked up by the face detector and their styles can be mixed with real people.
Fun fact: most of the faces used as inputs in the examples were actually generated from this app! Those people do not actually exist.
Note: This app does not collect any face data. All processing happens locally on the device. No data is shared with 3rd parties. No data is even stored locally. The only way to export data from this app is to share the generated image using the button in the lower right corner of the "stylized" image.
Some of the functionality of this app was adapted from the following research:
[1] Goodfellow, Ian, et al. "Generative adversarial nets." Advances in neural information processing systems 27 (2014).
[2] He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
[3] Karras, Tero, et al. "Analyzing and improving the image quality of stylegan." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.
[4] He, Kaiming, et al. "Momentum contrast for unsupervised visual representation learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.
[5] Huang, Yuge, et al. "Curricularface: adaptive curriculum learning loss for deep face recognition." proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.
[6] Richardson, Elad, et al. "Encoding in style: a stylegan encoder for image-to-image translation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.
[7] Tov, Omer, et al. "Designing an encoder for stylegan image manipulation." ACM Transactions on Graphics (TOG) 40.4 (2021): 1-14.
Additional Charges
The watermark can be removed and the generated image can be exported with an in app purchase. Each generated image requires its own in app purchase.
Here are a few ways this app can be used:
1) Hair color
Set the "input" photo to the person whose hair color you want to change. Choose a photo of a person with the hair color you want for the "style" photo. Since hair color is a relatively shallow style, styles can be mixed at shallow depths here. Mixing deeper styles will transfer deeper properties from the style photo such as face structure, gender, and head pose.
2) Baby generator
Choose photos of both parents and mix their styles. If mixing deep styles here, the gender styles will mix. There is a gender slider to adjust for this. If only mixing shallow styles, the gender will be that of the "input" face.
3) Hair transplant
This requires deeper style mixing than hair color. Set the input photo to the person who will receive the hair, and the style photo to someone who will donate the hair. Try to find the shallowest depth that transfers the hair style. This seems to be the 2nd to last and 3rd to last layers in most cases.
4) Cartoons
Some cartoon faces will be picked up by the face detector and their styles can be mixed with real people.
Fun fact: most of the faces used as inputs in the examples were actually generated from this app! Those people do not actually exist.
Note: This app does not collect any face data. All processing happens locally on the device. No data is shared with 3rd parties. No data is even stored locally. The only way to export data from this app is to share the generated image using the button in the lower right corner of the "stylized" image.
Some of the functionality of this app was adapted from the following research:
[1] Goodfellow, Ian, et al. "Generative adversarial nets." Advances in neural information processing systems 27 (2014).
[2] He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
[3] Karras, Tero, et al. "Analyzing and improving the image quality of stylegan." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.
[4] He, Kaiming, et al. "Momentum contrast for unsupervised visual representation learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.
[5] Huang, Yuge, et al. "Curricularface: adaptive curriculum learning loss for deep face recognition." proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.
[6] Richardson, Elad, et al. "Encoding in style: a stylegan encoder for image-to-image translation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.
[7] Tov, Omer, et al. "Designing an encoder for stylegan image manipulation." ACM Transactions on Graphics (TOG) 40.4 (2021): 1-14.
Show More
