AgingMapGAN (AMGAN): High-Resolution Controllable Face Aging with Spatially-Aware Conditional GANs
Video Summary
Abstract
Existing approaches and datasets for face aging produce results skewed towards the mean, with individual variations and expression wrinkles often invisible or overlooked in favor of global patterns such as the fattening of the face. Moreover, they offer little to no control over the way the faces are aged and can difficultly be scaled to large images, thus preventing their usage in many real-world applications. To address these limitations, we present an approach to change the appearance of a high-resolution image using ethnicity-specific aging information and weak spatial supervision to guide the aging process. We demonstrate the advantage of our proposed method in terms of quality, control, and how it can be used on high-definition images while limiting the computational overhead.
Paper & Supplementary Materials
AgingMapGAN (AMGAN): High-Resolution Controllable Face Aging with Spatially-Aware Conditional GANs.
ECCV, 2020 (AIM Workshop Oral) [arXiv] [BibTeX] [Supplementary Materials]
Model
Our model takes a patch p from the input image I, a target aging map A, and two orthonogal gradient images X and Y. The image patch Ip is then transformed according to the local aging information contained in the map Ap, while the orthogonal gradients Xp and Yp provide the coordinates of the patch in a fully-convolutional manner. The conditions are injected in the generator via the SPADE block to preserve the spatial information. Finally, the generator uses an attention mechanism to only change relevant parts of the image, thus preserving the clothes, earrings and other facial features unrelated to aging.
Training
To train our model, we use four different losses: LAge to penalize the aging map estimation, LLoc for the patch localization, LWGAN for the realism of the generated images, and LCyc for the fidelity to the original image.
Results
Supervised: High-Resolution Standardized Dataset
We have labeled 7000 images, using 15 clinical aging signs for each image, in order to build accurate ethnicity-specific aging maps for each individual. On this dataset, our model is able to generate aged and rejuvenated faces with complete control over the localization and amount of aging.
We recommend viewing the videos in full-screen to see the generated HD images (1024px).
We recommend opening the images in a new tab to see the details.
Weakly-Supervised: FFHQ
To train our model on the FFHQ dataset, we have created labels in a weakly-supervised fashion, using regression models trained on our labeled dataset. Despite this and the challenging poses, occlusions and lighting conditions of the dataset, our approach successfully ages and rejuvenates images in high-definition.
We recommend opening the images in a new tab to see the details.
Other works
Check out our other paper presented at AIM (ECCV 2020): https://robinkips.github.io/CA-GAN/